diff --git a/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/analytics/coremldata.bin b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..823d91cf23a2bb454a5b6dd99a9ae05da88006bc --- /dev/null +++ b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9880c37e22316efbaf285d84d1a4eecab31657806256346bd1939fcf4a775924 +size 243 diff --git a/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/coremldata.bin b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..6eed47e0dd8f2c62579405217a9755f463661f78 --- /dev/null +++ b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:914dd099519cff6e07d71c8a29081c50c800030c21d635e1a226ec7286819c8d +size 1292 diff --git a/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/metadata.json b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..3dc3df54de7517ae88aad148997bb18d28dc5797 --- /dev/null +++ b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND AMI streaming diarizer (pipeline, T=1, max_speakers=4)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 4)", + "shortDescription" : "", + "shape" : "[1, 1, 4]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 6 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 6, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 66, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 36, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 10, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 46 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 345)", + "shortDescription" : "", + "shape" : "[1, 1, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 6 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 6, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"ami\", \"model_label\": \"AMI\", \"variant\": \"pipeline\", \"chunk_size\": 1, \"step_duration_ms\": 100, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 4, \"max_nspks\": 6, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_ami_100ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/model.mil b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..69b8fb57d7fa52c2c479b44e6c51ee2bf32cccf4 --- /dev/null +++ b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/model.mil @@ -0,0 +1,1184 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0]])]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 1, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 1, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 1, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 1, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+0)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 1, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, x_3))[name = tensor("window_3")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = window_3)[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_249_split_sizes_0 = const()[name = tensor("op_249_split_sizes_0"), val = tensor([256, 256])]; + tensor var_249_axis_0 = const()[name = tensor("op_249_axis_0"), val = tensor(1)]; + tensor var_249_0, tensor var_249_1 = split(axis = var_249_axis_0, split_sizes = var_249_split_sizes_0, x = inputs_3)[name = tensor("op_249")]; + tensor var_251 = sigmoid(x = var_249_1)[name = tensor("op_251")]; + tensor inputs_5 = mul(x = var_249_0, y = var_251)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([1, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_282_begin_0 = const()[name = tensor("op_282_begin_0"), val = tensor([0, -1, 0])]; + tensor var_282_end_0 = const()[name = tensor("op_282_end_0"), val = tensor([1, 16, 256])]; + tensor var_282_end_mask_0 = const()[name = tensor("op_282_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_282 = slice_by_index(begin = var_282_begin_0, end = var_282_end_0, end_mask = var_282_end_mask_0, x = conv_out_1)[name = tensor("op_282")]; + tensor var_284_perm_0 = const()[name = tensor("op_284_perm_0"), val = tensor([1, 0, 2])]; + tensor var_284 = transpose(perm = var_284_perm_0, x = var_282)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_284)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_307 = const()[name = tensor("op_307"), val = tensor(0x1p-1)]; + tensor var_308 = mul(x = input_39, y = var_307)[name = tensor("op_308")]; + tensor input_41 = add(x = var_308, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_337 = const()[name = tensor("op_337"), val = tensor(0x1p-1)]; + tensor var_338 = mul(x = input_51, y = var_337)[name = tensor("op_338")]; + tensor input_53 = add(x = var_338, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_352 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_353 = const()[name = tensor("op_353"), val = tensor([1, 1, 4, 64])]; + tensor var_354 = reshape(shape = var_353, x = var_352)[name = tensor("op_354")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_358 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_359 = const()[name = tensor("op_359"), val = tensor(0x1p-3)]; + tensor var_360 = mul(x = var_358, y = var_359)[name = tensor("op_360")]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 1, 4, 64])]; + tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_366 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_367 = const()[name = tensor("op_367"), val = tensor([1, 1, 4, 64])]; + tensor var_368 = reshape(shape = var_367, x = var_366)[name = tensor("op_368")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_362)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_354)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 1])]; + tensor var_379 = reshape(shape = var_378, x = sqrt_s_t_3)[name = tensor("op_379")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_379)[name = tensor("M_3")]; + tensor var_381 = mul(x = qk_3, y = M_3)[name = tensor("op_381")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_368)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_381, y = v_3)[name = tensor("inner_3")]; + tensor var_383_transpose_x_0 = const()[name = tensor("op_383_transpose_x_0"), val = tensor(false)]; + tensor var_383_transpose_y_0 = const()[name = tensor("op_383_transpose_y_0"), val = tensor(false)]; + tensor var_383 = matmul(transpose_x = var_383_transpose_x_0, transpose_y = var_383_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_383")]; + tensor var_384 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_384")]; + tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 1, 1, 1])]; + tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; + tensor cross_3 = mul(x = var_383, y = var_386)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_389 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_389")]; + tensor var_391_transpose_x_1 = const()[name = tensor("op_391_transpose_x_1"), val = tensor(true)]; + tensor var_391_transpose_y_1 = const()[name = tensor("op_391_transpose_y_1"), val = tensor(false)]; + tensor var_391 = matmul(transpose_x = var_391_transpose_x_1, transpose_y = var_391_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_391")]; + tensor new_kv_unnorm_3 = add(x = var_389, y = var_391)[name = tensor("new_kv_unnorm_3")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor(0x1p+0)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_393)[name = tensor("new_scale_3")]; + tensor var_395 = sqrt(x = new_scale_3)[name = tensor("op_395")]; + tensor var_396 = real_div(x = new_kv_unnorm_3, y = var_395)[name = tensor("op_396")]; + tensor var_397_perm_0 = const()[name = tensor("op_397_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_397 = transpose(perm = var_397_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_397)[name = tensor("out_9")]; + tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 256])]; + tensor out_11 = reshape(shape = var_401, x = out_9)[name = tensor("out_11")]; + tensor var_403 = silu(x = input_57)[name = tensor("op_403")]; + tensor input_59 = mul(x = var_403, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_5_begin_0 = const()[name = tensor("window_5_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_5_end_0 = const()[name = tensor("window_5_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_5_end_mask_0 = const()[name = tensor("window_5_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_5_squeeze_mask_0 = const()[name = tensor("window_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_5 = slice_by_index(begin = window_5_begin_0, end = window_5_end_0, end_mask = window_5_end_mask_0, squeeze_mask = window_5_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_5")]; + tensor var_414_begin_0 = const()[name = tensor("op_414_begin_0"), val = tensor([0, 1, 0])]; + tensor var_414_end_0 = const()[name = tensor("op_414_end_0"), val = tensor([1, 16, 256])]; + tensor var_414_end_mask_0 = const()[name = tensor("op_414_end_mask_0"), val = tensor([true, true, true])]; + tensor var_414 = slice_by_index(begin = var_414_begin_0, end = var_414_end_0, end_mask = var_414_end_mask_0, x = window_5)[name = tensor("op_414")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_414, x_9))[name = tensor("window_7")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = window_7)[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_439_split_sizes_0 = const()[name = tensor("op_439_split_sizes_0"), val = tensor([256, 256])]; + tensor var_439_axis_0 = const()[name = tensor("op_439_axis_0"), val = tensor(1)]; + tensor var_439_0, tensor var_439_1 = split(axis = var_439_axis_0, split_sizes = var_439_split_sizes_0, x = inputs_13)[name = tensor("op_439")]; + tensor var_441 = sigmoid(x = var_439_1)[name = tensor("op_441")]; + tensor inputs_15 = mul(x = var_439_0, y = var_441)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([1, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_472_begin_0 = const()[name = tensor("op_472_begin_0"), val = tensor([0, -1, 0])]; + tensor var_472_end_0 = const()[name = tensor("op_472_end_0"), val = tensor([1, 16, 256])]; + tensor var_472_end_mask_0 = const()[name = tensor("op_472_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_472 = slice_by_index(begin = var_472_begin_0, end = var_472_end_0, end_mask = var_472_end_mask_0, x = conv_out_3)[name = tensor("op_472")]; + tensor var_474_perm_0 = const()[name = tensor("op_474_perm_0"), val = tensor([1, 0, 2])]; + tensor var_474 = transpose(perm = var_474_perm_0, x = var_472)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_474)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_497 = const()[name = tensor("op_497"), val = tensor(0x1p-1)]; + tensor var_498 = mul(x = input_79, y = var_497)[name = tensor("op_498")]; + tensor input_81 = add(x = var_498, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_527 = const()[name = tensor("op_527"), val = tensor(0x1p-1)]; + tensor var_528 = mul(x = input_91, y = var_527)[name = tensor("op_528")]; + tensor input_93 = add(x = var_528, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_542 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 1, 4, 64])]; + tensor var_544 = reshape(shape = var_543, x = var_542)[name = tensor("op_544")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_548 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_549 = const()[name = tensor("op_549"), val = tensor(0x1p-3)]; + tensor var_550 = mul(x = var_548, y = var_549)[name = tensor("op_550")]; + tensor var_551 = const()[name = tensor("op_551"), val = tensor([1, 1, 4, 64])]; + tensor var_552 = reshape(shape = var_551, x = var_550)[name = tensor("op_552")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_556 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_557 = const()[name = tensor("op_557"), val = tensor([1, 1, 4, 64])]; + tensor var_558 = reshape(shape = var_557, x = var_556)[name = tensor("op_558")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_552)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_544)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_568 = const()[name = tensor("op_568"), val = tensor([1, 1])]; + tensor var_569 = reshape(shape = var_568, x = sqrt_s_t_5)[name = tensor("op_569")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_569)[name = tensor("M_5")]; + tensor var_571 = mul(x = qk_5, y = M_5)[name = tensor("op_571")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_558)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_571, y = v_5)[name = tensor("inner_5")]; + tensor var_573_transpose_x_0 = const()[name = tensor("op_573_transpose_x_0"), val = tensor(false)]; + tensor var_573_transpose_y_0 = const()[name = tensor("op_573_transpose_y_0"), val = tensor(false)]; + tensor var_573 = matmul(transpose_x = var_573_transpose_x_0, transpose_y = var_573_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_573")]; + tensor var_574 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_574")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 1, 1, 1])]; + tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; + tensor cross_5 = mul(x = var_573, y = var_576)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_579 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_579")]; + tensor var_581_transpose_x_1 = const()[name = tensor("op_581_transpose_x_1"), val = tensor(true)]; + tensor var_581_transpose_y_1 = const()[name = tensor("op_581_transpose_y_1"), val = tensor(false)]; + tensor var_581 = matmul(transpose_x = var_581_transpose_x_1, transpose_y = var_581_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_581")]; + tensor new_kv_unnorm_5 = add(x = var_579, y = var_581)[name = tensor("new_kv_unnorm_5")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor(0x1p+0)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_583)[name = tensor("new_scale_5")]; + tensor var_585 = sqrt(x = new_scale_5)[name = tensor("op_585")]; + tensor var_586 = real_div(x = new_kv_unnorm_5, y = var_585)[name = tensor("op_586")]; + tensor var_587_perm_0 = const()[name = tensor("op_587_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_587 = transpose(perm = var_587_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_587)[name = tensor("out_15")]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 256])]; + tensor out_17 = reshape(shape = var_591, x = out_15)[name = tensor("out_17")]; + tensor var_593 = silu(x = input_97)[name = tensor("op_593")]; + tensor input_99 = mul(x = var_593, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; + tensor var_604_begin_0 = const()[name = tensor("op_604_begin_0"), val = tensor([0, 1, 0])]; + tensor var_604_end_0 = const()[name = tensor("op_604_end_0"), val = tensor([1, 16, 256])]; + tensor var_604_end_mask_0 = const()[name = tensor("op_604_end_mask_0"), val = tensor([true, true, true])]; + tensor var_604 = slice_by_index(begin = var_604_begin_0, end = var_604_end_0, end_mask = var_604_end_mask_0, x = window_9)[name = tensor("op_604")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_604, x_15))[name = tensor("window_11")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = window_11)[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_629_split_sizes_0 = const()[name = tensor("op_629_split_sizes_0"), val = tensor([256, 256])]; + tensor var_629_axis_0 = const()[name = tensor("op_629_axis_0"), val = tensor(1)]; + tensor var_629_0, tensor var_629_1 = split(axis = var_629_axis_0, split_sizes = var_629_split_sizes_0, x = inputs_23)[name = tensor("op_629")]; + tensor var_631 = sigmoid(x = var_629_1)[name = tensor("op_631")]; + tensor inputs_25 = mul(x = var_629_0, y = var_631)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([1, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_662_begin_0 = const()[name = tensor("op_662_begin_0"), val = tensor([0, -1, 0])]; + tensor var_662_end_0 = const()[name = tensor("op_662_end_0"), val = tensor([1, 16, 256])]; + tensor var_662_end_mask_0 = const()[name = tensor("op_662_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_662 = slice_by_index(begin = var_662_begin_0, end = var_662_end_0, end_mask = var_662_end_mask_0, x = conv_out_5)[name = tensor("op_662")]; + tensor var_664_perm_0 = const()[name = tensor("op_664_perm_0"), val = tensor([1, 0, 2])]; + tensor var_664 = transpose(perm = var_664_perm_0, x = var_662)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_664)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_687 = const()[name = tensor("op_687"), val = tensor(0x1p-1)]; + tensor var_688 = mul(x = input_119, y = var_687)[name = tensor("op_688")]; + tensor input_121 = add(x = var_688, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_717 = const()[name = tensor("op_717"), val = tensor(0x1p-1)]; + tensor var_718 = mul(x = input_131, y = var_717)[name = tensor("op_718")]; + tensor input_133 = add(x = var_718, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_732 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_733 = const()[name = tensor("op_733"), val = tensor([1, 1, 4, 64])]; + tensor var_734 = reshape(shape = var_733, x = var_732)[name = tensor("op_734")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_738 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_739 = const()[name = tensor("op_739"), val = tensor(0x1p-3)]; + tensor var_740 = mul(x = var_738, y = var_739)[name = tensor("op_740")]; + tensor var_741 = const()[name = tensor("op_741"), val = tensor([1, 1, 4, 64])]; + tensor var_742 = reshape(shape = var_741, x = var_740)[name = tensor("op_742")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_746 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 1, 4, 64])]; + tensor var_748 = reshape(shape = var_747, x = var_746)[name = tensor("op_748")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_742)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_734)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_758 = const()[name = tensor("op_758"), val = tensor([1, 1])]; + tensor var_759 = reshape(shape = var_758, x = sqrt_s_t_7)[name = tensor("op_759")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_759)[name = tensor("M_7")]; + tensor var_761 = mul(x = qk_7, y = M_7)[name = tensor("op_761")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_748)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_761, y = v_7)[name = tensor("inner_7")]; + tensor var_763_transpose_x_0 = const()[name = tensor("op_763_transpose_x_0"), val = tensor(false)]; + tensor var_763_transpose_y_0 = const()[name = tensor("op_763_transpose_y_0"), val = tensor(false)]; + tensor var_763 = matmul(transpose_x = var_763_transpose_x_0, transpose_y = var_763_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_763")]; + tensor var_764 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_764")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 1, 1, 1])]; + tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; + tensor cross_7 = mul(x = var_763, y = var_766)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_769 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_769")]; + tensor var_771_transpose_x_1 = const()[name = tensor("op_771_transpose_x_1"), val = tensor(true)]; + tensor var_771_transpose_y_1 = const()[name = tensor("op_771_transpose_y_1"), val = tensor(false)]; + tensor var_771 = matmul(transpose_x = var_771_transpose_x_1, transpose_y = var_771_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_771")]; + tensor new_kv_unnorm_7 = add(x = var_769, y = var_771)[name = tensor("new_kv_unnorm_7")]; + tensor var_773 = const()[name = tensor("op_773"), val = tensor(0x1p+0)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_773)[name = tensor("new_scale_7")]; + tensor var_775 = sqrt(x = new_scale_7)[name = tensor("op_775")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_775)[name = tensor("nkv_1")]; + tensor var_777_perm_0 = const()[name = tensor("op_777_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_777 = transpose(perm = var_777_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_777)[name = tensor("out_21")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 1, 256])]; + tensor out_23 = reshape(shape = var_781, x = out_21)[name = tensor("out_23")]; + tensor var_783 = silu(x = input_137)[name = tensor("op_783")]; + tensor input_139 = mul(x = var_783, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; + tensor var_794_begin_0 = const()[name = tensor("op_794_begin_0"), val = tensor([0, 1, 0])]; + tensor var_794_end_0 = const()[name = tensor("op_794_end_0"), val = tensor([1, 16, 256])]; + tensor var_794_end_mask_0 = const()[name = tensor("op_794_end_mask_0"), val = tensor([true, true, true])]; + tensor var_794 = slice_by_index(begin = var_794_begin_0, end = var_794_end_0, end_mask = var_794_end_mask_0, x = window_13)[name = tensor("op_794")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_794, x_21))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = window)[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_819_split_sizes_0 = const()[name = tensor("op_819_split_sizes_0"), val = tensor([256, 256])]; + tensor var_819_axis_0 = const()[name = tensor("op_819_axis_0"), val = tensor(1)]; + tensor var_819_0, tensor var_819_1 = split(axis = var_819_axis_0, split_sizes = var_819_split_sizes_0, x = inputs_33)[name = tensor("op_819")]; + tensor var_821 = sigmoid(x = var_819_1)[name = tensor("op_821")]; + tensor inputs_35 = mul(x = var_819_0, y = var_821)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([1, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_852_begin_0 = const()[name = tensor("op_852_begin_0"), val = tensor([0, -1, 0])]; + tensor var_852_end_0 = const()[name = tensor("op_852_end_0"), val = tensor([1, 16, 256])]; + tensor var_852_end_mask_0 = const()[name = tensor("op_852_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_852 = slice_by_index(begin = var_852_begin_0, end = var_852_end_0, end_mask = var_852_end_mask_0, x = conv_out_7)[name = tensor("op_852")]; + tensor var_854_perm_0 = const()[name = tensor("op_854_perm_0"), val = tensor([1, 0, 2])]; + tensor var_854 = transpose(perm = var_854_perm_0, x = var_852)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_854)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_877 = const()[name = tensor("op_877"), val = tensor(0x1p-1)]; + tensor var_878 = mul(x = input_159, y = var_877)[name = tensor("op_878")]; + tensor input_161 = add(x = var_878, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 0, 1])]; + tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 256, 19])]; + tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = cat)[name = tensor("op_896")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_898 = const()[name = tensor("op_898"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_899 = reduce_l2_norm(axes = var_898, keep_dims = var_29, x = input_163)[name = tensor("op_899")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_899)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_903_axis_0 = const()[name = tensor("op_903_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_903_axis_0, values = (var_206, var_396, var_586, nkv_1))[name = tensor("op_903")]; + tensor var_905_axis_0 = const()[name = tensor("op_905_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_905_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_905")]; + tensor var_907_axis_0 = const()[name = tensor("op_907_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_907_axis_0, values = (window_3, window_7, window_11, window))[name = tensor("op_907")]; + tensor var_916 = const()[name = tensor("op_916"), val = tensor(0x1.5798eep-27)]; + tensor var_921 = const()[name = tensor("op_921"), val = tensor(0x1.4f8b58p-17)]; + tensor var_923 = const()[name = tensor("op_923"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_924 = const()[name = tensor("op_924"), val = tensor(true)]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor(0x1p+0)]; + tensor var_930 = const()[name = tensor("op_930"), val = tensor(-1)]; + tensor var_936 = const()[name = tensor("op_936"), val = tensor(0)]; + tensor var_993 = const()[name = tensor("op_993"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_998_axes_0 = const()[name = tensor("op_998_axes_0"), val = tensor([2])]; + tensor var_998 = expand_dims(axes = var_998_axes_0, x = emb)[name = tensor("op_998")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 6, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_998)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_930, interleave = input_165_interleave_0, values = (emb_exp, var_993))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1006_perm_0 = const()[name = tensor("op_1006_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([6, 1, 256])]; + tensor var_1006 = transpose(perm = var_1006_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1010, x = var_1006)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 6, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1018 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1019 = const()[name = tensor("op_1019"), val = tensor([6, 1, 4, 64])]; + tensor var_1020 = reshape(shape = var_1019, x = var_1018)[name = tensor("op_1020")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1024 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1025 = const()[name = tensor("op_1025"), val = tensor(0x1p-3)]; + tensor var_1026 = mul(x = var_1024, y = var_1025)[name = tensor("op_1026")]; + tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([6, 1, 4, 64])]; + tensor var_1028 = reshape(shape = var_1027, x = var_1026)[name = tensor("op_1028")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1032 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1033 = const()[name = tensor("op_1033"), val = tensor([6, 1, 4, 64])]; + tensor var_1034 = reshape(shape = var_1033, x = var_1032)[name = tensor("op_1034")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_936, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_926, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1028)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1020)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 1])]; + tensor var_1047 = reshape(shape = var_1046, x = valid_mask)[name = tensor("op_1047")]; + tensor var_1049 = const()[name = tensor("op_1049"), val = tensor([1, 1])]; + tensor var_1050 = reshape(shape = var_1049, x = sqrt_s_t_9)[name = tensor("op_1050")]; + tensor M_9 = real_div(x = var_1047, y = var_1050)[name = tensor("M_9")]; + tensor var_1052 = mul(x = qk_9, y = M_9)[name = tensor("op_1052")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1034)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1052, y = v_9)[name = tensor("inner_9")]; + tensor var_1054_transpose_x_0 = const()[name = tensor("op_1054_transpose_x_0"), val = tensor(false)]; + tensor var_1054_transpose_y_0 = const()[name = tensor("op_1054_transpose_y_0"), val = tensor(false)]; + tensor var_1054 = matmul(transpose_x = var_1054_transpose_x_0, transpose_y = var_1054_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1054")]; + tensor var_1055 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1055")]; + tensor var_1056 = const()[name = tensor("op_1056"), val = tensor([1, 1, 1, 1])]; + tensor var_1057 = reshape(shape = var_1056, x = var_1055)[name = tensor("op_1057")]; + tensor cross_9 = mul(x = var_1054, y = var_1057)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1060 = const()[name = tensor("op_1060"), val = tensor([1, 1, 1, 1])]; + tensor var_1061 = reshape(shape = var_1060, x = valid_mask)[name = tensor("op_1061")]; + tensor v_masked_1 = mul(x = v_9, y = var_1061)[name = tensor("v_masked_1")]; + tensor var_1063 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1063")]; + tensor var_1065_transpose_x_1 = const()[name = tensor("op_1065_transpose_x_1"), val = tensor(true)]; + tensor var_1065_transpose_y_1 = const()[name = tensor("op_1065_transpose_y_1"), val = tensor(false)]; + tensor var_1065 = matmul(transpose_x = var_1065_transpose_x_1, transpose_y = var_1065_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1065")]; + tensor new_kv_unnorm_9 = add(x = var_1063, y = var_1065)[name = tensor("new_kv_unnorm_9")]; + tensor var_1067_keep_dims_0 = const()[name = tensor("op_1067_keep_dims_0"), val = tensor(false)]; + tensor var_1067 = reduce_sum(keep_dims = var_1067_keep_dims_0, x = valid_mask)[name = tensor("op_1067")]; + tensor var_1068 = const()[name = tensor("op_1068"), val = tensor([1])]; + tensor var_1069 = reshape(shape = var_1068, x = var_1067)[name = tensor("op_1069")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1069)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_926, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1073 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1073")]; + tensor var_1074_perm_0 = const()[name = tensor("op_1074_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1074 = transpose(perm = var_1074_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_923, x = var_1074)[name = tensor("out_27")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([6, 1, 256])]; + tensor out_29 = reshape(shape = var_1078, x = out_27)[name = tensor("out_29")]; + tensor var_1080 = silu(x = input_169)[name = tensor("op_1080")]; + tensor input_171 = mul(x = var_1080, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_921, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1090 = const()[name = tensor("op_1090"), val = tensor([1, 6, 1, 256])]; + tensor var_1091 = reshape(shape = var_1090, x = xt_1)[name = tensor("op_1091")]; + tensor var_1092_perm_0 = const()[name = tensor("op_1092_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1095 = const()[name = tensor("op_1095"), val = tensor([1, 6, 256])]; + tensor var_1092 = transpose(perm = var_1092_perm_0, x = var_1091)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1095, x = var_1092)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1118 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([6, 1, 3, 256])]; + tensor var_1120 = reshape(shape = concat_1, x = var_1118)[name = tensor("op_1120")]; + tensor var_1121_axes_0 = const()[name = tensor("op_1121_axes_0"), val = tensor([0])]; + tensor var_1121 = expand_dims(axes = var_1121_axes_0, x = var_1120)[name = tensor("op_1121")]; + tensor var_1122_perm_0 = const()[name = tensor("op_1122_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1123_axes_0 = const()[name = tensor("op_1123_axes_0"), val = tensor([-2])]; + tensor var_1122 = transpose(perm = var_1122_perm_0, x = var_1121)[name = tensor("transpose_21")]; + tensor var_1123 = squeeze(axes = var_1123_axes_0, x = var_1122)[name = tensor("op_1123")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 6, 1, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1123)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 6, 1, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1123)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 6, 1, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1123)[name = tensor("v_11")]; + tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([6, 4, 64])]; + tensor var_1132 = reshape(shape = var_1131, x = q_11)[name = tensor("op_1132")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1138 = const()[name = tensor("op_1138"), val = tensor([6, 4, 64])]; + tensor var_1139 = reshape(shape = var_1138, x = k_11)[name = tensor("op_1139")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([6, 4, 64])]; + tensor var_1146 = reshape(shape = var_1145, x = v_11)[name = tensor("op_1146")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1149 = const()[name = tensor("op_1149"), val = tensor([1, 4, 6, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1132)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1149, x = q_13)[name = tensor("q_15")]; + tensor var_1151 = const()[name = tensor("op_1151"), val = tensor([1, 4, 6, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1139)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1151, x = k_13)[name = tensor("k_15")]; + tensor var_1153 = const()[name = tensor("op_1153"), val = tensor([1, 4, 6, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1146)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1153, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([2, 0, 1, 3])]; + tensor var_1161 = const()[name = tensor("op_1161"), val = tensor([6, 256])]; + tensor var_1157 = transpose(perm = var_1156, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1161, x = var_1157)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1165 = const()[name = tensor("op_1165"), val = tensor([6, 1, 256])]; + tensor attn_output_7 = reshape(shape = var_1165, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_921, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_921, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 6, 256])]; + tensor x_31 = reshape(shape = var_1185, x = xt_3)[name = tensor("x_31")]; + tensor var_1187_perm_0 = const()[name = tensor("op_1187_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([6, 1, 256])]; + tensor var_1187 = transpose(perm = var_1187_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1191, x = var_1187)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 6, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1199 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1200 = const()[name = tensor("op_1200"), val = tensor([6, 1, 4, 64])]; + tensor var_1201 = reshape(shape = var_1200, x = var_1199)[name = tensor("op_1201")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1205 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1206 = const()[name = tensor("op_1206"), val = tensor(0x1p-3)]; + tensor var_1207 = mul(x = var_1205, y = var_1206)[name = tensor("op_1207")]; + tensor var_1208 = const()[name = tensor("op_1208"), val = tensor([6, 1, 4, 64])]; + tensor var_1209 = reshape(shape = var_1208, x = var_1207)[name = tensor("op_1209")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1213 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1214 = const()[name = tensor("op_1214"), val = tensor([6, 1, 4, 64])]; + tensor var_1215 = reshape(shape = var_1214, x = var_1213)[name = tensor("op_1215")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_926, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1209)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1201)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 1])]; + tensor var_1231 = reshape(shape = var_1230, x = sqrt_s_t)[name = tensor("op_1231")]; + tensor M = real_div(x = var_1047, y = var_1231)[name = tensor("M")]; + tensor var_1233 = mul(x = qk, y = M)[name = tensor("op_1233")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1215)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1233, y = v_17)[name = tensor("inner")]; + tensor var_1235_transpose_x_0 = const()[name = tensor("op_1235_transpose_x_0"), val = tensor(false)]; + tensor var_1235_transpose_y_0 = const()[name = tensor("op_1235_transpose_y_0"), val = tensor(false)]; + tensor var_1235 = matmul(transpose_x = var_1235_transpose_x_0, transpose_y = var_1235_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1235")]; + tensor var_1236 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1236")]; + tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1, 1, 1])]; + tensor var_1238 = reshape(shape = var_1237, x = var_1236)[name = tensor("op_1238")]; + tensor cross = mul(x = var_1235, y = var_1238)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1061)[name = tensor("v_masked")]; + tensor var_1244 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1244")]; + tensor var_1246_transpose_x_1 = const()[name = tensor("op_1246_transpose_x_1"), val = tensor(true)]; + tensor var_1246_transpose_y_1 = const()[name = tensor("op_1246_transpose_y_1"), val = tensor(false)]; + tensor var_1246 = matmul(transpose_x = var_1246_transpose_x_1, transpose_y = var_1246_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1246")]; + tensor new_kv_unnorm = add(x = var_1244, y = var_1246)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1069)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_926, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1255_perm_0 = const()[name = tensor("op_1255_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1255 = transpose(perm = var_1255_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_923, x = var_1255)[name = tensor("out_33")]; + tensor var_1259 = const()[name = tensor("op_1259"), val = tensor([6, 1, 256])]; + tensor out = reshape(shape = var_1259, x = out_33)[name = tensor("out")]; + tensor var_1261 = silu(x = input_187)[name = tensor("op_1261")]; + tensor input_189 = mul(x = var_1261, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_921, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1271 = const()[name = tensor("op_1271"), val = tensor([1, 6, 1, 256])]; + tensor var_1272 = reshape(shape = var_1271, x = xt_5)[name = tensor("op_1272")]; + tensor var_1273_perm_0 = const()[name = tensor("op_1273_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1276 = const()[name = tensor("op_1276"), val = tensor([1, 6, 256])]; + tensor var_1273 = transpose(perm = var_1273_perm_0, x = var_1272)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1276, x = var_1273)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1299 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([6, 1, 3, 256])]; + tensor var_1301 = reshape(shape = concat_2, x = var_1299)[name = tensor("op_1301")]; + tensor var_1302_axes_0 = const()[name = tensor("op_1302_axes_0"), val = tensor([0])]; + tensor var_1302 = expand_dims(axes = var_1302_axes_0, x = var_1301)[name = tensor("op_1302")]; + tensor var_1303_perm_0 = const()[name = tensor("op_1303_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1304_axes_0 = const()[name = tensor("op_1304_axes_0"), val = tensor([-2])]; + tensor var_1303 = transpose(perm = var_1303_perm_0, x = var_1302)[name = tensor("transpose_8")]; + tensor var_1304 = squeeze(axes = var_1304_axes_0, x = var_1303)[name = tensor("op_1304")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 6, 1, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1304)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 6, 1, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1304)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 6, 1, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1304)[name = tensor("v_19")]; + tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([6, 4, 64])]; + tensor var_1313 = reshape(shape = var_1312, x = q_19)[name = tensor("op_1313")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1319 = const()[name = tensor("op_1319"), val = tensor([6, 4, 64])]; + tensor var_1320 = reshape(shape = var_1319, x = k_19)[name = tensor("op_1320")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([6, 4, 64])]; + tensor var_1327 = reshape(shape = var_1326, x = v_19)[name = tensor("op_1327")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1330 = const()[name = tensor("op_1330"), val = tensor([1, 4, 6, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1313)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1330, x = q_21)[name = tensor("q")]; + tensor var_1332 = const()[name = tensor("op_1332"), val = tensor([1, 4, 6, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1320)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1332, x = k_21)[name = tensor("k")]; + tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 4, 6, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1327)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1334, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([2, 0, 1, 3])]; + tensor var_1342 = const()[name = tensor("op_1342"), val = tensor([6, 256])]; + tensor var_1338 = transpose(perm = var_1337, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1342, x = var_1338)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1346 = const()[name = tensor("op_1346"), val = tensor([6, 1, 256])]; + tensor attn_output = reshape(shape = var_1346, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_921, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_921, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 6, 256])]; + tensor input = reshape(shape = var_1366, x = xt)[name = tensor("input")]; + tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([-1])]; + tensor var_1369 = reduce_l2_norm(axes = var_1368, keep_dims = var_924, x = input)[name = tensor("op_1369")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_916, beta = const_42, x = var_1369)[name = tensor("clip_5")]; + tensor var_1371 = real_div(x = input, y = clip_5)[name = tensor("op_1371")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([1, 256, 6])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1371)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = emb, y = reshape_1)[name = tensor("matmul_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 1, 5])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = matmul_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1375")]; + tensor var_1377_axis_0 = const()[name = tensor("op_1377_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1377_axis_0, values = (var_1073, nkv))[name = tensor("op_1377")]; + tensor var_1379_axis_0 = const()[name = tensor("op_1379_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1379_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1379")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/weights/weight.bin b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..42066c1f3eae2a9c06856cf1737a94c0ad89d5b1 --- /dev/null +++ b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f77a11836cfdeff2a24070c9eb01c8a3146d23216ebd63a5f02bf154a5e95aa7 +size 44401536 diff --git a/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..23b9a3151223c06a795711a72944fec0a2d73987 --- /dev/null +++ b/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d99472f0bd00e091613210906f7fb7be42e92f6019968246af665c07dab02a7 +size 171348 diff --git a/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..42066c1f3eae2a9c06856cf1737a94c0ad89d5b1 --- /dev/null +++ b/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f77a11836cfdeff2a24070c9eb01c8a3146d23216ebd63a5f02bf154a5e95aa7 +size 44401536 diff --git a/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Manifest.json b/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..6a6df0d6218423e5cf5828262039e34f982ae069 --- /dev/null +++ b/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "25D17A51-8136-403A-8E7C-65E586D88F62": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "966E268D-F501-43F0-BD33-968439381A4D": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "966E268D-F501-43F0-BD33-968439381A4D" +} diff --git a/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/analytics/coremldata.bin b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..dbf2843795d71fc2c8081b28675d1065af2fc06c --- /dev/null +++ b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6a81fd874a82f12a3d9273121dfc5cf2c7b7fc3ab75940416ce700d6c464715 +size 243 diff --git a/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/coremldata.bin b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c2e0a156964f1117ec1437b8f2af288e3cae4b6a --- /dev/null +++ b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ff7a04daff876566a5951933f1afdcf6a2bc73cf8214ddcb95845fb83a7ad66 +size 1292 diff --git a/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/metadata.json b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..fceb4211edaa544fa740da0dbf819fbec3b8ddf1 --- /dev/null +++ b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND AMI streaming diarizer (pipeline, T=2, max_speakers=4)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 2 × 4)", + "shortDescription" : "", + "shape" : "[1, 2, 4]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 6 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 6, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 68, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 48, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 14, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 47 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 2 × 345)", + "shortDescription" : "", + "shape" : "[1, 2, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 6 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 6, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2)", + "shortDescription" : "", + "shape" : "[2]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"ami\", \"model_label\": \"AMI\", \"variant\": \"pipeline\", \"chunk_size\": 2, \"step_duration_ms\": 200, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 4, \"max_nspks\": 6, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_ami_200ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/model.mil b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..06fd0af3c049963c87924d0249246fe8af7890b0 --- /dev/null +++ b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/model.mil @@ -0,0 +1,1246 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 2, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 2, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 2, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([2, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 2, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 2, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; + tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; + tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; + tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; + tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; + tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 1, 256])]; + tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, true, true])]; + tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; + tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; + tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; + tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; + tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5))[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_257_split_sizes_0 = const()[name = tensor("op_257_split_sizes_0"), val = tensor([256, 256])]; + tensor var_257_axis_0 = const()[name = tensor("op_257_axis_0"), val = tensor(1)]; + tensor var_257_0, tensor var_257_1 = split(axis = var_257_axis_0, split_sizes = var_257_split_sizes_0, x = inputs_3)[name = tensor("op_257")]; + tensor var_259 = sigmoid(x = var_257_1)[name = tensor("op_259")]; + tensor inputs_5 = mul(x = var_257_0, y = var_259)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([2, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_290_begin_0 = const()[name = tensor("op_290_begin_0"), val = tensor([0, -1, 0])]; + tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([2, 16, 256])]; + tensor var_290_end_mask_0 = const()[name = tensor("op_290_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_290 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, x = conv_out_1)[name = tensor("op_290")]; + tensor var_292_perm_0 = const()[name = tensor("op_292_perm_0"), val = tensor([1, 0, 2])]; + tensor var_292 = transpose(perm = var_292_perm_0, x = var_290)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_292)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_315 = const()[name = tensor("op_315"), val = tensor(0x1p-1)]; + tensor var_316 = mul(x = input_39, y = var_315)[name = tensor("op_316")]; + tensor input_41 = add(x = var_316, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_345 = const()[name = tensor("op_345"), val = tensor(0x1p-1)]; + tensor var_346 = mul(x = input_51, y = var_345)[name = tensor("op_346")]; + tensor input_53 = add(x = var_346, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_360 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 2, 4, 64])]; + tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_366 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_367 = const()[name = tensor("op_367"), val = tensor(0x1p-3)]; + tensor var_368 = mul(x = var_366, y = var_367)[name = tensor("op_368")]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 2, 4, 64])]; + tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_374 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_375 = const()[name = tensor("op_375"), val = tensor([1, 2, 4, 64])]; + tensor var_376 = reshape(shape = var_375, x = var_374)[name = tensor("op_376")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_370)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_362)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([2, 1])]; + tensor var_387 = reshape(shape = var_386, x = sqrt_s_t_3)[name = tensor("op_387")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_387)[name = tensor("M_3")]; + tensor var_389 = mul(x = qk_3, y = M_3)[name = tensor("op_389")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_376)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_389, y = v_3)[name = tensor("inner_3")]; + tensor var_391_transpose_x_0 = const()[name = tensor("op_391_transpose_x_0"), val = tensor(false)]; + tensor var_391_transpose_y_0 = const()[name = tensor("op_391_transpose_y_0"), val = tensor(false)]; + tensor var_391 = matmul(transpose_x = var_391_transpose_x_0, transpose_y = var_391_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_391")]; + tensor var_392 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_392")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor([1, 1, 2, 1])]; + tensor var_394 = reshape(shape = var_393, x = var_392)[name = tensor("op_394")]; + tensor cross_3 = mul(x = var_391, y = var_394)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_397 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_397")]; + tensor var_399_transpose_x_1 = const()[name = tensor("op_399_transpose_x_1"), val = tensor(true)]; + tensor var_399_transpose_y_1 = const()[name = tensor("op_399_transpose_y_1"), val = tensor(false)]; + tensor var_399 = matmul(transpose_x = var_399_transpose_x_1, transpose_y = var_399_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_399")]; + tensor new_kv_unnorm_3 = add(x = var_397, y = var_399)[name = tensor("new_kv_unnorm_3")]; + tensor var_401 = const()[name = tensor("op_401"), val = tensor(0x1p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_401)[name = tensor("new_scale_3")]; + tensor var_403 = sqrt(x = new_scale_3)[name = tensor("op_403")]; + tensor var_404 = real_div(x = new_kv_unnorm_3, y = var_403)[name = tensor("op_404")]; + tensor var_405_perm_0 = const()[name = tensor("op_405_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_405 = transpose(perm = var_405_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_405)[name = tensor("out_9")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 2, 256])]; + tensor out_11 = reshape(shape = var_409, x = out_9)[name = tensor("out_11")]; + tensor var_411 = silu(x = input_57)[name = tensor("op_411")]; + tensor input_59 = mul(x = var_411, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_7_begin_0 = const()[name = tensor("window_7_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_7_end_0 = const()[name = tensor("window_7_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_7_end_mask_0 = const()[name = tensor("window_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_7_squeeze_mask_0 = const()[name = tensor("window_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_7 = slice_by_index(begin = window_7_begin_0, end = window_7_end_0, end_mask = window_7_end_mask_0, squeeze_mask = window_7_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_7")]; + tensor var_419_begin_0 = const()[name = tensor("op_419_begin_0"), val = tensor([0, 0, 0])]; + tensor var_419_end_0 = const()[name = tensor("op_419_end_0"), val = tensor([1, 1, 256])]; + tensor var_419_end_mask_0 = const()[name = tensor("op_419_end_mask_0"), val = tensor([true, false, true])]; + tensor var_419 = slice_by_index(begin = var_419_begin_0, end = var_419_end_0, end_mask = var_419_end_mask_0, x = x_9)[name = tensor("op_419")]; + tensor var_422_begin_0 = const()[name = tensor("op_422_begin_0"), val = tensor([0, 1, 0])]; + tensor var_422_end_0 = const()[name = tensor("op_422_end_0"), val = tensor([1, 16, 256])]; + tensor var_422_end_mask_0 = const()[name = tensor("op_422_end_mask_0"), val = tensor([true, true, true])]; + tensor var_422 = slice_by_index(begin = var_422_begin_0, end = var_422_end_0, end_mask = var_422_end_mask_0, x = window_7)[name = tensor("op_422")]; + tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; + tensor window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_422, var_419))[name = tensor("window_9")]; + tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 1, 0])]; + tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; + tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, true, true])]; + tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; + tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; + tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; + tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; + tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_9, window_11))[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_455_split_sizes_0 = const()[name = tensor("op_455_split_sizes_0"), val = tensor([256, 256])]; + tensor var_455_axis_0 = const()[name = tensor("op_455_axis_0"), val = tensor(1)]; + tensor var_455_0, tensor var_455_1 = split(axis = var_455_axis_0, split_sizes = var_455_split_sizes_0, x = inputs_13)[name = tensor("op_455")]; + tensor var_457 = sigmoid(x = var_455_1)[name = tensor("op_457")]; + tensor inputs_15 = mul(x = var_455_0, y = var_457)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([2, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_488_begin_0 = const()[name = tensor("op_488_begin_0"), val = tensor([0, -1, 0])]; + tensor var_488_end_0 = const()[name = tensor("op_488_end_0"), val = tensor([2, 16, 256])]; + tensor var_488_end_mask_0 = const()[name = tensor("op_488_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_488 = slice_by_index(begin = var_488_begin_0, end = var_488_end_0, end_mask = var_488_end_mask_0, x = conv_out_3)[name = tensor("op_488")]; + tensor var_490_perm_0 = const()[name = tensor("op_490_perm_0"), val = tensor([1, 0, 2])]; + tensor var_490 = transpose(perm = var_490_perm_0, x = var_488)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_490)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_513 = const()[name = tensor("op_513"), val = tensor(0x1p-1)]; + tensor var_514 = mul(x = input_79, y = var_513)[name = tensor("op_514")]; + tensor input_81 = add(x = var_514, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_543 = const()[name = tensor("op_543"), val = tensor(0x1p-1)]; + tensor var_544 = mul(x = input_91, y = var_543)[name = tensor("op_544")]; + tensor input_93 = add(x = var_544, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_558 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_559 = const()[name = tensor("op_559"), val = tensor([1, 2, 4, 64])]; + tensor var_560 = reshape(shape = var_559, x = var_558)[name = tensor("op_560")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_564 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_565 = const()[name = tensor("op_565"), val = tensor(0x1p-3)]; + tensor var_566 = mul(x = var_564, y = var_565)[name = tensor("op_566")]; + tensor var_567 = const()[name = tensor("op_567"), val = tensor([1, 2, 4, 64])]; + tensor var_568 = reshape(shape = var_567, x = var_566)[name = tensor("op_568")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_572 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 2, 4, 64])]; + tensor var_574 = reshape(shape = var_573, x = var_572)[name = tensor("op_574")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_568)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_560)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_584 = const()[name = tensor("op_584"), val = tensor([2, 1])]; + tensor var_585 = reshape(shape = var_584, x = sqrt_s_t_5)[name = tensor("op_585")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_585)[name = tensor("M_5")]; + tensor var_587 = mul(x = qk_5, y = M_5)[name = tensor("op_587")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_574)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_587, y = v_5)[name = tensor("inner_5")]; + tensor var_589_transpose_x_0 = const()[name = tensor("op_589_transpose_x_0"), val = tensor(false)]; + tensor var_589_transpose_y_0 = const()[name = tensor("op_589_transpose_y_0"), val = tensor(false)]; + tensor var_589 = matmul(transpose_x = var_589_transpose_x_0, transpose_y = var_589_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_589")]; + tensor var_590 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_590")]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 2, 1])]; + tensor var_592 = reshape(shape = var_591, x = var_590)[name = tensor("op_592")]; + tensor cross_5 = mul(x = var_589, y = var_592)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_595 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_595")]; + tensor var_597_transpose_x_1 = const()[name = tensor("op_597_transpose_x_1"), val = tensor(true)]; + tensor var_597_transpose_y_1 = const()[name = tensor("op_597_transpose_y_1"), val = tensor(false)]; + tensor var_597 = matmul(transpose_x = var_597_transpose_x_1, transpose_y = var_597_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_597")]; + tensor new_kv_unnorm_5 = add(x = var_595, y = var_597)[name = tensor("new_kv_unnorm_5")]; + tensor var_599 = const()[name = tensor("op_599"), val = tensor(0x1p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_599)[name = tensor("new_scale_5")]; + tensor var_601 = sqrt(x = new_scale_5)[name = tensor("op_601")]; + tensor var_602 = real_div(x = new_kv_unnorm_5, y = var_601)[name = tensor("op_602")]; + tensor var_603_perm_0 = const()[name = tensor("op_603_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_603 = transpose(perm = var_603_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_603)[name = tensor("out_15")]; + tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 2, 256])]; + tensor out_17 = reshape(shape = var_607, x = out_15)[name = tensor("out_17")]; + tensor var_609 = silu(x = input_97)[name = tensor("op_609")]; + tensor input_99 = mul(x = var_609, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; + tensor var_617_begin_0 = const()[name = tensor("op_617_begin_0"), val = tensor([0, 0, 0])]; + tensor var_617_end_0 = const()[name = tensor("op_617_end_0"), val = tensor([1, 1, 256])]; + tensor var_617_end_mask_0 = const()[name = tensor("op_617_end_mask_0"), val = tensor([true, false, true])]; + tensor var_617 = slice_by_index(begin = var_617_begin_0, end = var_617_end_0, end_mask = var_617_end_mask_0, x = x_15)[name = tensor("op_617")]; + tensor var_620_begin_0 = const()[name = tensor("op_620_begin_0"), val = tensor([0, 1, 0])]; + tensor var_620_end_0 = const()[name = tensor("op_620_end_0"), val = tensor([1, 16, 256])]; + tensor var_620_end_mask_0 = const()[name = tensor("op_620_end_mask_0"), val = tensor([true, true, true])]; + tensor var_620 = slice_by_index(begin = var_620_begin_0, end = var_620_end_0, end_mask = var_620_end_mask_0, x = window_13)[name = tensor("op_620")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_620, var_617))[name = tensor("window_15")]; + tensor var_625_begin_0 = const()[name = tensor("op_625_begin_0"), val = tensor([0, 1, 0])]; + tensor var_625_end_0 = const()[name = tensor("op_625_end_0"), val = tensor([1, 1, 256])]; + tensor var_625_end_mask_0 = const()[name = tensor("op_625_end_mask_0"), val = tensor([true, true, true])]; + tensor var_625 = slice_by_index(begin = var_625_begin_0, end = var_625_end_0, end_mask = var_625_end_mask_0, x = x_15)[name = tensor("op_625")]; + tensor var_628_begin_0 = const()[name = tensor("op_628_begin_0"), val = tensor([0, 1, 0])]; + tensor var_628_end_0 = const()[name = tensor("op_628_end_0"), val = tensor([1, 16, 256])]; + tensor var_628_end_mask_0 = const()[name = tensor("op_628_end_mask_0"), val = tensor([true, true, true])]; + tensor var_628 = slice_by_index(begin = var_628_begin_0, end = var_628_end_0, end_mask = var_628_end_mask_0, x = window_15)[name = tensor("op_628")]; + tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; + tensor window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_628, var_625))[name = tensor("window_17")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_15, window_17))[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_653_split_sizes_0 = const()[name = tensor("op_653_split_sizes_0"), val = tensor([256, 256])]; + tensor var_653_axis_0 = const()[name = tensor("op_653_axis_0"), val = tensor(1)]; + tensor var_653_0, tensor var_653_1 = split(axis = var_653_axis_0, split_sizes = var_653_split_sizes_0, x = inputs_23)[name = tensor("op_653")]; + tensor var_655 = sigmoid(x = var_653_1)[name = tensor("op_655")]; + tensor inputs_25 = mul(x = var_653_0, y = var_655)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([2, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_686_begin_0 = const()[name = tensor("op_686_begin_0"), val = tensor([0, -1, 0])]; + tensor var_686_end_0 = const()[name = tensor("op_686_end_0"), val = tensor([2, 16, 256])]; + tensor var_686_end_mask_0 = const()[name = tensor("op_686_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_686 = slice_by_index(begin = var_686_begin_0, end = var_686_end_0, end_mask = var_686_end_mask_0, x = conv_out_5)[name = tensor("op_686")]; + tensor var_688_perm_0 = const()[name = tensor("op_688_perm_0"), val = tensor([1, 0, 2])]; + tensor var_688 = transpose(perm = var_688_perm_0, x = var_686)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_688)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_711 = const()[name = tensor("op_711"), val = tensor(0x1p-1)]; + tensor var_712 = mul(x = input_119, y = var_711)[name = tensor("op_712")]; + tensor input_121 = add(x = var_712, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_741 = const()[name = tensor("op_741"), val = tensor(0x1p-1)]; + tensor var_742 = mul(x = input_131, y = var_741)[name = tensor("op_742")]; + tensor input_133 = add(x = var_742, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_756 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_757 = const()[name = tensor("op_757"), val = tensor([1, 2, 4, 64])]; + tensor var_758 = reshape(shape = var_757, x = var_756)[name = tensor("op_758")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_762 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_763 = const()[name = tensor("op_763"), val = tensor(0x1p-3)]; + tensor var_764 = mul(x = var_762, y = var_763)[name = tensor("op_764")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 2, 4, 64])]; + tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_770 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 2, 4, 64])]; + tensor var_772 = reshape(shape = var_771, x = var_770)[name = tensor("op_772")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_766)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_758)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_782 = const()[name = tensor("op_782"), val = tensor([2, 1])]; + tensor var_783 = reshape(shape = var_782, x = sqrt_s_t_7)[name = tensor("op_783")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_783)[name = tensor("M_7")]; + tensor var_785 = mul(x = qk_7, y = M_7)[name = tensor("op_785")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_772)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_785, y = v_7)[name = tensor("inner_7")]; + tensor var_787_transpose_x_0 = const()[name = tensor("op_787_transpose_x_0"), val = tensor(false)]; + tensor var_787_transpose_y_0 = const()[name = tensor("op_787_transpose_y_0"), val = tensor(false)]; + tensor var_787 = matmul(transpose_x = var_787_transpose_x_0, transpose_y = var_787_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_787")]; + tensor var_788 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_788")]; + tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 1, 2, 1])]; + tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; + tensor cross_7 = mul(x = var_787, y = var_790)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_793 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_793")]; + tensor var_795_transpose_x_1 = const()[name = tensor("op_795_transpose_x_1"), val = tensor(true)]; + tensor var_795_transpose_y_1 = const()[name = tensor("op_795_transpose_y_1"), val = tensor(false)]; + tensor var_795 = matmul(transpose_x = var_795_transpose_x_1, transpose_y = var_795_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_795")]; + tensor new_kv_unnorm_7 = add(x = var_793, y = var_795)[name = tensor("new_kv_unnorm_7")]; + tensor var_797 = const()[name = tensor("op_797"), val = tensor(0x1p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_797)[name = tensor("new_scale_7")]; + tensor var_799 = sqrt(x = new_scale_7)[name = tensor("op_799")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_799)[name = tensor("nkv_1")]; + tensor var_801_perm_0 = const()[name = tensor("op_801_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_801 = transpose(perm = var_801_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_801)[name = tensor("out_21")]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 2, 256])]; + tensor out_23 = reshape(shape = var_805, x = out_21)[name = tensor("out_23")]; + tensor var_807 = silu(x = input_137)[name = tensor("op_807")]; + tensor input_139 = mul(x = var_807, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_19_begin_0 = const()[name = tensor("window_19_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_19_end_0 = const()[name = tensor("window_19_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_19_end_mask_0 = const()[name = tensor("window_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_19_squeeze_mask_0 = const()[name = tensor("window_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_19 = slice_by_index(begin = window_19_begin_0, end = window_19_end_0, end_mask = window_19_end_mask_0, squeeze_mask = window_19_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_19")]; + tensor var_815_begin_0 = const()[name = tensor("op_815_begin_0"), val = tensor([0, 0, 0])]; + tensor var_815_end_0 = const()[name = tensor("op_815_end_0"), val = tensor([1, 1, 256])]; + tensor var_815_end_mask_0 = const()[name = tensor("op_815_end_mask_0"), val = tensor([true, false, true])]; + tensor var_815 = slice_by_index(begin = var_815_begin_0, end = var_815_end_0, end_mask = var_815_end_mask_0, x = x_21)[name = tensor("op_815")]; + tensor var_818_begin_0 = const()[name = tensor("op_818_begin_0"), val = tensor([0, 1, 0])]; + tensor var_818_end_0 = const()[name = tensor("op_818_end_0"), val = tensor([1, 16, 256])]; + tensor var_818_end_mask_0 = const()[name = tensor("op_818_end_mask_0"), val = tensor([true, true, true])]; + tensor var_818 = slice_by_index(begin = var_818_begin_0, end = var_818_end_0, end_mask = var_818_end_mask_0, x = window_19)[name = tensor("op_818")]; + tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; + tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_818, var_815))[name = tensor("window_21")]; + tensor var_823_begin_0 = const()[name = tensor("op_823_begin_0"), val = tensor([0, 1, 0])]; + tensor var_823_end_0 = const()[name = tensor("op_823_end_0"), val = tensor([1, 1, 256])]; + tensor var_823_end_mask_0 = const()[name = tensor("op_823_end_mask_0"), val = tensor([true, true, true])]; + tensor var_823 = slice_by_index(begin = var_823_begin_0, end = var_823_end_0, end_mask = var_823_end_mask_0, x = x_21)[name = tensor("op_823")]; + tensor var_826_begin_0 = const()[name = tensor("op_826_begin_0"), val = tensor([0, 1, 0])]; + tensor var_826_end_0 = const()[name = tensor("op_826_end_0"), val = tensor([1, 16, 256])]; + tensor var_826_end_mask_0 = const()[name = tensor("op_826_end_mask_0"), val = tensor([true, true, true])]; + tensor var_826 = slice_by_index(begin = var_826_begin_0, end = var_826_end_0, end_mask = var_826_end_mask_0, x = window_21)[name = tensor("op_826")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_826, var_823))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_21, window))[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_851_split_sizes_0 = const()[name = tensor("op_851_split_sizes_0"), val = tensor([256, 256])]; + tensor var_851_axis_0 = const()[name = tensor("op_851_axis_0"), val = tensor(1)]; + tensor var_851_0, tensor var_851_1 = split(axis = var_851_axis_0, split_sizes = var_851_split_sizes_0, x = inputs_33)[name = tensor("op_851")]; + tensor var_853 = sigmoid(x = var_851_1)[name = tensor("op_853")]; + tensor inputs_35 = mul(x = var_851_0, y = var_853)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([2, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_884_begin_0 = const()[name = tensor("op_884_begin_0"), val = tensor([0, -1, 0])]; + tensor var_884_end_0 = const()[name = tensor("op_884_end_0"), val = tensor([2, 16, 256])]; + tensor var_884_end_mask_0 = const()[name = tensor("op_884_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_884 = slice_by_index(begin = var_884_begin_0, end = var_884_end_0, end_mask = var_884_end_mask_0, x = conv_out_7)[name = tensor("op_884")]; + tensor var_886_perm_0 = const()[name = tensor("op_886_perm_0"), val = tensor([1, 0, 2])]; + tensor var_886 = transpose(perm = var_886_perm_0, x = var_884)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_886)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_909 = const()[name = tensor("op_909"), val = tensor(0x1p-1)]; + tensor var_910 = mul(x = input_159, y = var_909)[name = tensor("op_910")]; + tensor input_161 = add(x = var_910, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_928_begin_0 = const()[name = tensor("op_928_begin_0"), val = tensor([0, 0, 2])]; + tensor var_928_end_0 = const()[name = tensor("op_928_end_0"), val = tensor([1, 256, 20])]; + tensor var_928_end_mask_0 = const()[name = tensor("op_928_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_928_begin_0, end = var_928_end_0, end_mask = var_928_end_mask_0, x = cat)[name = tensor("op_928")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_930 = const()[name = tensor("op_930"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_931 = reduce_l2_norm(axes = var_930, keep_dims = var_29, x = input_163)[name = tensor("op_931")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_931)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_935_axis_0 = const()[name = tensor("op_935_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_935_axis_0, values = (var_206, var_404, var_602, nkv_1))[name = tensor("op_935")]; + tensor var_937_axis_0 = const()[name = tensor("op_937_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_937_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_937")]; + tensor var_939_axis_0 = const()[name = tensor("op_939_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_939_axis_0, values = (window_5, window_11, window_17, window))[name = tensor("op_939")]; + tensor var_948 = const()[name = tensor("op_948"), val = tensor(0x1.5798eep-27)]; + tensor var_953 = const()[name = tensor("op_953"), val = tensor(0x1.4f8b58p-17)]; + tensor var_955 = const()[name = tensor("op_955"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_956 = const()[name = tensor("op_956"), val = tensor(true)]; + tensor var_958 = const()[name = tensor("op_958"), val = tensor(0x1p+0)]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor(-1)]; + tensor var_968 = const()[name = tensor("op_968"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_1030_axes_0 = const()[name = tensor("op_1030_axes_0"), val = tensor([2])]; + tensor var_1030 = expand_dims(axes = var_1030_axes_0, x = emb)[name = tensor("op_1030")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 6, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1030)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_962, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1038_perm_0 = const()[name = tensor("op_1038_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1042 = const()[name = tensor("op_1042"), val = tensor([6, 2, 256])]; + tensor var_1038 = transpose(perm = var_1038_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1042, x = var_1038)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 6, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1050 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1051 = const()[name = tensor("op_1051"), val = tensor([6, 2, 4, 64])]; + tensor var_1052 = reshape(shape = var_1051, x = var_1050)[name = tensor("op_1052")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1056 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1057 = const()[name = tensor("op_1057"), val = tensor(0x1p-3)]; + tensor var_1058 = mul(x = var_1056, y = var_1057)[name = tensor("op_1058")]; + tensor var_1059 = const()[name = tensor("op_1059"), val = tensor([6, 2, 4, 64])]; + tensor var_1060 = reshape(shape = var_1059, x = var_1058)[name = tensor("op_1060")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1064 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1065 = const()[name = tensor("op_1065"), val = tensor([6, 2, 4, 64])]; + tensor var_1066 = reshape(shape = var_1065, x = var_1064)[name = tensor("op_1066")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_968, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_958, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1060)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1052)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([1, 2])]; + tensor var_1079 = reshape(shape = var_1078, x = valid_mask)[name = tensor("op_1079")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1079)[name = tensor("causal_with_valid_1")]; + tensor var_1081 = const()[name = tensor("op_1081"), val = tensor([2, 1])]; + tensor var_1082 = reshape(shape = var_1081, x = sqrt_s_t_9)[name = tensor("op_1082")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1082)[name = tensor("M_9")]; + tensor var_1084 = mul(x = qk_9, y = M_9)[name = tensor("op_1084")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1066)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1084, y = v_9)[name = tensor("inner_9")]; + tensor var_1086_transpose_x_0 = const()[name = tensor("op_1086_transpose_x_0"), val = tensor(false)]; + tensor var_1086_transpose_y_0 = const()[name = tensor("op_1086_transpose_y_0"), val = tensor(false)]; + tensor var_1086 = matmul(transpose_x = var_1086_transpose_x_0, transpose_y = var_1086_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1086")]; + tensor var_1087 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1087")]; + tensor var_1088 = const()[name = tensor("op_1088"), val = tensor([1, 1, 2, 1])]; + tensor var_1089 = reshape(shape = var_1088, x = var_1087)[name = tensor("op_1089")]; + tensor cross_9 = mul(x = var_1086, y = var_1089)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([1, 1, 2, 1])]; + tensor var_1093 = reshape(shape = var_1092, x = valid_mask)[name = tensor("op_1093")]; + tensor v_masked_1 = mul(x = v_9, y = var_1093)[name = tensor("v_masked_1")]; + tensor var_1095 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1095")]; + tensor var_1097_transpose_x_1 = const()[name = tensor("op_1097_transpose_x_1"), val = tensor(true)]; + tensor var_1097_transpose_y_1 = const()[name = tensor("op_1097_transpose_y_1"), val = tensor(false)]; + tensor var_1097 = matmul(transpose_x = var_1097_transpose_x_1, transpose_y = var_1097_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1097")]; + tensor new_kv_unnorm_9 = add(x = var_1095, y = var_1097)[name = tensor("new_kv_unnorm_9")]; + tensor var_1099_keep_dims_0 = const()[name = tensor("op_1099_keep_dims_0"), val = tensor(false)]; + tensor var_1099 = reduce_sum(keep_dims = var_1099_keep_dims_0, x = valid_mask)[name = tensor("op_1099")]; + tensor var_1100 = const()[name = tensor("op_1100"), val = tensor([1])]; + tensor var_1101 = reshape(shape = var_1100, x = var_1099)[name = tensor("op_1101")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1101)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_958, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1105 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1105")]; + tensor var_1106_perm_0 = const()[name = tensor("op_1106_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1106 = transpose(perm = var_1106_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_955, x = var_1106)[name = tensor("out_27")]; + tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([6, 2, 256])]; + tensor out_29 = reshape(shape = var_1110, x = out_27)[name = tensor("out_29")]; + tensor var_1112 = silu(x = input_169)[name = tensor("op_1112")]; + tensor input_171 = mul(x = var_1112, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_953, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1122 = const()[name = tensor("op_1122"), val = tensor([1, 6, 2, 256])]; + tensor var_1123 = reshape(shape = var_1122, x = xt_1)[name = tensor("op_1123")]; + tensor var_1124_perm_0 = const()[name = tensor("op_1124_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([2, 6, 256])]; + tensor var_1124 = transpose(perm = var_1124_perm_0, x = var_1123)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1127, x = var_1124)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1150 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([6, 2, 3, 256])]; + tensor var_1152 = reshape(shape = concat_1, x = var_1150)[name = tensor("op_1152")]; + tensor var_1153_axes_0 = const()[name = tensor("op_1153_axes_0"), val = tensor([0])]; + tensor var_1153 = expand_dims(axes = var_1153_axes_0, x = var_1152)[name = tensor("op_1153")]; + tensor var_1154_perm_0 = const()[name = tensor("op_1154_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1155_axes_0 = const()[name = tensor("op_1155_axes_0"), val = tensor([-2])]; + tensor var_1154 = transpose(perm = var_1154_perm_0, x = var_1153)[name = tensor("transpose_21")]; + tensor var_1155 = squeeze(axes = var_1155_axes_0, x = var_1154)[name = tensor("op_1155")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 6, 2, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1155)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 6, 2, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1155)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 6, 2, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1155)[name = tensor("v_11")]; + tensor var_1163 = const()[name = tensor("op_1163"), val = tensor([6, 8, 64])]; + tensor var_1164 = reshape(shape = var_1163, x = q_11)[name = tensor("op_1164")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1170 = const()[name = tensor("op_1170"), val = tensor([6, 8, 64])]; + tensor var_1171 = reshape(shape = var_1170, x = k_11)[name = tensor("op_1171")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor([6, 8, 64])]; + tensor var_1178 = reshape(shape = var_1177, x = v_11)[name = tensor("op_1178")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1181 = const()[name = tensor("op_1181"), val = tensor([2, 4, 6, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1164)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1181, x = q_13)[name = tensor("q_15")]; + tensor var_1183 = const()[name = tensor("op_1183"), val = tensor([2, 4, 6, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1171)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1183, x = k_13)[name = tensor("k_15")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([2, 4, 6, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1178)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1185, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1188 = const()[name = tensor("op_1188"), val = tensor([2, 0, 1, 3])]; + tensor var_1193 = const()[name = tensor("op_1193"), val = tensor([12, 256])]; + tensor var_1189 = transpose(perm = var_1188, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1193, x = var_1189)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([6, 2, 256])]; + tensor attn_output_7 = reshape(shape = var_1197, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_953, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_953, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([1, 2, 6, 256])]; + tensor x_31 = reshape(shape = var_1217, x = xt_3)[name = tensor("x_31")]; + tensor var_1219_perm_0 = const()[name = tensor("op_1219_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1223 = const()[name = tensor("op_1223"), val = tensor([6, 2, 256])]; + tensor var_1219 = transpose(perm = var_1219_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1223, x = var_1219)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 6, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1231 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([6, 2, 4, 64])]; + tensor var_1233 = reshape(shape = var_1232, x = var_1231)[name = tensor("op_1233")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1237 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor(0x1p-3)]; + tensor var_1239 = mul(x = var_1237, y = var_1238)[name = tensor("op_1239")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([6, 2, 4, 64])]; + tensor var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor("op_1241")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1245 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1246 = const()[name = tensor("op_1246"), val = tensor([6, 2, 4, 64])]; + tensor var_1247 = reshape(shape = var_1246, x = var_1245)[name = tensor("op_1247")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_958, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1241)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1233)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([2, 1])]; + tensor var_1263 = reshape(shape = var_1262, x = sqrt_s_t)[name = tensor("op_1263")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1263)[name = tensor("M")]; + tensor var_1265 = mul(x = qk, y = M)[name = tensor("op_1265")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1247)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1265, y = v_17)[name = tensor("inner")]; + tensor var_1267_transpose_x_0 = const()[name = tensor("op_1267_transpose_x_0"), val = tensor(false)]; + tensor var_1267_transpose_y_0 = const()[name = tensor("op_1267_transpose_y_0"), val = tensor(false)]; + tensor var_1267 = matmul(transpose_x = var_1267_transpose_x_0, transpose_y = var_1267_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1267")]; + tensor var_1268 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1268")]; + tensor var_1269 = const()[name = tensor("op_1269"), val = tensor([1, 1, 2, 1])]; + tensor var_1270 = reshape(shape = var_1269, x = var_1268)[name = tensor("op_1270")]; + tensor cross = mul(x = var_1267, y = var_1270)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1093)[name = tensor("v_masked")]; + tensor var_1276 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1276")]; + tensor var_1278_transpose_x_1 = const()[name = tensor("op_1278_transpose_x_1"), val = tensor(true)]; + tensor var_1278_transpose_y_1 = const()[name = tensor("op_1278_transpose_y_1"), val = tensor(false)]; + tensor var_1278 = matmul(transpose_x = var_1278_transpose_x_1, transpose_y = var_1278_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1278")]; + tensor new_kv_unnorm = add(x = var_1276, y = var_1278)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1101)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_958, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1287_perm_0 = const()[name = tensor("op_1287_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1287 = transpose(perm = var_1287_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_955, x = var_1287)[name = tensor("out_33")]; + tensor var_1291 = const()[name = tensor("op_1291"), val = tensor([6, 2, 256])]; + tensor out = reshape(shape = var_1291, x = out_33)[name = tensor("out")]; + tensor var_1293 = silu(x = input_187)[name = tensor("op_1293")]; + tensor input_189 = mul(x = var_1293, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_953, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1303 = const()[name = tensor("op_1303"), val = tensor([1, 6, 2, 256])]; + tensor var_1304 = reshape(shape = var_1303, x = xt_5)[name = tensor("op_1304")]; + tensor var_1305_perm_0 = const()[name = tensor("op_1305_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1308 = const()[name = tensor("op_1308"), val = tensor([2, 6, 256])]; + tensor var_1305 = transpose(perm = var_1305_perm_0, x = var_1304)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1308, x = var_1305)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1331 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([6, 2, 3, 256])]; + tensor var_1333 = reshape(shape = concat_2, x = var_1331)[name = tensor("op_1333")]; + tensor var_1334_axes_0 = const()[name = tensor("op_1334_axes_0"), val = tensor([0])]; + tensor var_1334 = expand_dims(axes = var_1334_axes_0, x = var_1333)[name = tensor("op_1334")]; + tensor var_1335_perm_0 = const()[name = tensor("op_1335_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1336_axes_0 = const()[name = tensor("op_1336_axes_0"), val = tensor([-2])]; + tensor var_1335 = transpose(perm = var_1335_perm_0, x = var_1334)[name = tensor("transpose_8")]; + tensor var_1336 = squeeze(axes = var_1336_axes_0, x = var_1335)[name = tensor("op_1336")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 6, 2, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1336)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 6, 2, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1336)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 6, 2, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1336)[name = tensor("v_19")]; + tensor var_1344 = const()[name = tensor("op_1344"), val = tensor([6, 8, 64])]; + tensor var_1345 = reshape(shape = var_1344, x = q_19)[name = tensor("op_1345")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1351 = const()[name = tensor("op_1351"), val = tensor([6, 8, 64])]; + tensor var_1352 = reshape(shape = var_1351, x = k_19)[name = tensor("op_1352")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([6, 8, 64])]; + tensor var_1359 = reshape(shape = var_1358, x = v_19)[name = tensor("op_1359")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1362 = const()[name = tensor("op_1362"), val = tensor([2, 4, 6, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1345)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1362, x = q_21)[name = tensor("q")]; + tensor var_1364 = const()[name = tensor("op_1364"), val = tensor([2, 4, 6, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1352)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1364, x = k_21)[name = tensor("k")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([2, 4, 6, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1359)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1366, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([2, 0, 1, 3])]; + tensor var_1374 = const()[name = tensor("op_1374"), val = tensor([12, 256])]; + tensor var_1370 = transpose(perm = var_1369, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1374, x = var_1370)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1378 = const()[name = tensor("op_1378"), val = tensor([6, 2, 256])]; + tensor attn_output = reshape(shape = var_1378, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_953, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_953, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([1, 2, 6, 256])]; + tensor input = reshape(shape = var_1398, x = xt)[name = tensor("input")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([-1])]; + tensor var_1401 = reduce_l2_norm(axes = var_1400, keep_dims = var_956, x = input)[name = tensor("op_1401")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_948, beta = const_42, x = var_1401)[name = tensor("clip_5")]; + tensor var_1403 = real_div(x = input, y = clip_5)[name = tensor("op_1403")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([2, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([2, 256, 6])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1403)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 2, 6])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 2, 5])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1407")]; + tensor var_1409_axis_0 = const()[name = tensor("op_1409_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1409_axis_0, values = (var_1105, nkv))[name = tensor("op_1409")]; + tensor var_1411_axis_0 = const()[name = tensor("op_1411_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1411_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1411")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/weights/weight.bin b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..4fd00b388372bd41848cd3ae1520261fa9f6835f --- /dev/null +++ b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42117736f66f329731a7688c50d40cd5cee0b5dc72a310b56ed7ce606bf6eac8 +size 44407680 diff --git a/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..4c678f86789d8217225a774ec066ff29f1c6adcc --- /dev/null +++ b/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:286832ef1477a85722d0e808565e32c2edf2c7abbbb9f2b3447253ac30f39155 +size 179858 diff --git a/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..4fd00b388372bd41848cd3ae1520261fa9f6835f --- /dev/null +++ b/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42117736f66f329731a7688c50d40cd5cee0b5dc72a310b56ed7ce606bf6eac8 +size 44407680 diff --git a/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Manifest.json b/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..612bcaf8e0b488f282381fd935e727601819bea6 --- /dev/null +++ b/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "9AA855BB-15E6-4440-985F-9D01D6A86BD7": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "FE3BD696-8DA2-4220-BA67-6DFD3C1F41A7": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "FE3BD696-8DA2-4220-BA67-6DFD3C1F41A7" +} diff --git a/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/analytics/coremldata.bin b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..3f6639b57a269384ee54f9f1660607c52c504864 --- /dev/null +++ b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6f56029ce7301e97b0cefc55ff6f7d7d8a47604b531cac5a18717b8934c6c12 +size 243 diff --git a/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/coremldata.bin b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..0adcb3383c298e39175a387aa8c5ed93b77e17d8 --- /dev/null +++ b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5272855c6f2aed3afbb9d4479de3da636f7780e628ff7ad02dc5bbc5f5d7894 +size 1292 diff --git a/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/metadata.json b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..24752739dc034358ad51ac236bb8ea00423b379e --- /dev/null +++ b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND AMI streaming diarizer (pipeline, T=3, max_speakers=4)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 3 × 4)", + "shortDescription" : "", + "shape" : "[1, 3, 4]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 6 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 6, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 68, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 56, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 18, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 47 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 3 × 345)", + "shortDescription" : "", + "shape" : "[1, 3, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 6 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 6, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 3)", + "shortDescription" : "", + "shape" : "[3]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"ami\", \"model_label\": \"AMI\", \"variant\": \"pipeline\", \"chunk_size\": 3, \"step_duration_ms\": 300, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 4, \"max_nspks\": 6, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_ami_300ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/model.mil b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..970596a9f4827af95d9232b6600321f0fda4a1c0 --- /dev/null +++ b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/model.mil @@ -0,0 +1,1286 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 3, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 3, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 3, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([3, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 3, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1.8p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 3, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; + tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; + tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; + tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; + tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; + tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; + tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; + tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; + tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; + tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; + tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; + tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; + tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; + tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 1, 256])]; + tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, true, true])]; + tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; + tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; + tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; + tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; + tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7))[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_265_split_sizes_0 = const()[name = tensor("op_265_split_sizes_0"), val = tensor([256, 256])]; + tensor var_265_axis_0 = const()[name = tensor("op_265_axis_0"), val = tensor(1)]; + tensor var_265_0, tensor var_265_1 = split(axis = var_265_axis_0, split_sizes = var_265_split_sizes_0, x = inputs_3)[name = tensor("op_265")]; + tensor var_267 = sigmoid(x = var_265_1)[name = tensor("op_267")]; + tensor inputs_5 = mul(x = var_265_0, y = var_267)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([3, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_298_begin_0 = const()[name = tensor("op_298_begin_0"), val = tensor([0, -1, 0])]; + tensor var_298_end_0 = const()[name = tensor("op_298_end_0"), val = tensor([3, 16, 256])]; + tensor var_298_end_mask_0 = const()[name = tensor("op_298_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_298 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, x = conv_out_1)[name = tensor("op_298")]; + tensor var_300_perm_0 = const()[name = tensor("op_300_perm_0"), val = tensor([1, 0, 2])]; + tensor var_300 = transpose(perm = var_300_perm_0, x = var_298)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_300)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_323 = const()[name = tensor("op_323"), val = tensor(0x1p-1)]; + tensor var_324 = mul(x = input_39, y = var_323)[name = tensor("op_324")]; + tensor input_41 = add(x = var_324, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_353 = const()[name = tensor("op_353"), val = tensor(0x1p-1)]; + tensor var_354 = mul(x = input_51, y = var_353)[name = tensor("op_354")]; + tensor input_53 = add(x = var_354, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_368 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 3, 4, 64])]; + tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_374 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_375 = const()[name = tensor("op_375"), val = tensor(0x1p-3)]; + tensor var_376 = mul(x = var_374, y = var_375)[name = tensor("op_376")]; + tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 3, 4, 64])]; + tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_382 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_383 = const()[name = tensor("op_383"), val = tensor([1, 3, 4, 64])]; + tensor var_384 = reshape(shape = var_383, x = var_382)[name = tensor("op_384")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_378)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_370)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_394 = const()[name = tensor("op_394"), val = tensor([3, 1])]; + tensor var_395 = reshape(shape = var_394, x = sqrt_s_t_3)[name = tensor("op_395")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_395)[name = tensor("M_3")]; + tensor var_397 = mul(x = qk_3, y = M_3)[name = tensor("op_397")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_384)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_397, y = v_3)[name = tensor("inner_3")]; + tensor var_399_transpose_x_0 = const()[name = tensor("op_399_transpose_x_0"), val = tensor(false)]; + tensor var_399_transpose_y_0 = const()[name = tensor("op_399_transpose_y_0"), val = tensor(false)]; + tensor var_399 = matmul(transpose_x = var_399_transpose_x_0, transpose_y = var_399_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_399")]; + tensor var_400 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_400")]; + tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 3, 1])]; + tensor var_402 = reshape(shape = var_401, x = var_400)[name = tensor("op_402")]; + tensor cross_3 = mul(x = var_399, y = var_402)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_405 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_405")]; + tensor var_407_transpose_x_1 = const()[name = tensor("op_407_transpose_x_1"), val = tensor(true)]; + tensor var_407_transpose_y_1 = const()[name = tensor("op_407_transpose_y_1"), val = tensor(false)]; + tensor var_407 = matmul(transpose_x = var_407_transpose_x_1, transpose_y = var_407_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_407")]; + tensor new_kv_unnorm_3 = add(x = var_405, y = var_407)[name = tensor("new_kv_unnorm_3")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor(0x1.8p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_409)[name = tensor("new_scale_3")]; + tensor var_411 = sqrt(x = new_scale_3)[name = tensor("op_411")]; + tensor var_412 = real_div(x = new_kv_unnorm_3, y = var_411)[name = tensor("op_412")]; + tensor var_413_perm_0 = const()[name = tensor("op_413_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_413 = transpose(perm = var_413_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_413)[name = tensor("out_9")]; + tensor var_417 = const()[name = tensor("op_417"), val = tensor([1, 3, 256])]; + tensor out_11 = reshape(shape = var_417, x = out_9)[name = tensor("out_11")]; + tensor var_419 = silu(x = input_57)[name = tensor("op_419")]; + tensor input_59 = mul(x = var_419, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; + tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 0, 0])]; + tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; + tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, false, true])]; + tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; + tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; + tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; + tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; + tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; + tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 1, 0])]; + tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 2, 256])]; + tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; + tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; + tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; + tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; + tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; + tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; + tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; + tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 2, 0])]; + tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 1, 256])]; + tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, true, true])]; + tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; + tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; + tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; + tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; + tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_11, window_13, window_15))[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_471_split_sizes_0 = const()[name = tensor("op_471_split_sizes_0"), val = tensor([256, 256])]; + tensor var_471_axis_0 = const()[name = tensor("op_471_axis_0"), val = tensor(1)]; + tensor var_471_0, tensor var_471_1 = split(axis = var_471_axis_0, split_sizes = var_471_split_sizes_0, x = inputs_13)[name = tensor("op_471")]; + tensor var_473 = sigmoid(x = var_471_1)[name = tensor("op_473")]; + tensor inputs_15 = mul(x = var_471_0, y = var_473)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([3, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_504_begin_0 = const()[name = tensor("op_504_begin_0"), val = tensor([0, -1, 0])]; + tensor var_504_end_0 = const()[name = tensor("op_504_end_0"), val = tensor([3, 16, 256])]; + tensor var_504_end_mask_0 = const()[name = tensor("op_504_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_504 = slice_by_index(begin = var_504_begin_0, end = var_504_end_0, end_mask = var_504_end_mask_0, x = conv_out_3)[name = tensor("op_504")]; + tensor var_506_perm_0 = const()[name = tensor("op_506_perm_0"), val = tensor([1, 0, 2])]; + tensor var_506 = transpose(perm = var_506_perm_0, x = var_504)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_506)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_529 = const()[name = tensor("op_529"), val = tensor(0x1p-1)]; + tensor var_530 = mul(x = input_79, y = var_529)[name = tensor("op_530")]; + tensor input_81 = add(x = var_530, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_559 = const()[name = tensor("op_559"), val = tensor(0x1p-1)]; + tensor var_560 = mul(x = input_91, y = var_559)[name = tensor("op_560")]; + tensor input_93 = add(x = var_560, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_574 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 3, 4, 64])]; + tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_580 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_581 = const()[name = tensor("op_581"), val = tensor(0x1p-3)]; + tensor var_582 = mul(x = var_580, y = var_581)[name = tensor("op_582")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 3, 4, 64])]; + tensor var_584 = reshape(shape = var_583, x = var_582)[name = tensor("op_584")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_588 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_589 = const()[name = tensor("op_589"), val = tensor([1, 3, 4, 64])]; + tensor var_590 = reshape(shape = var_589, x = var_588)[name = tensor("op_590")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_584)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_576)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_600 = const()[name = tensor("op_600"), val = tensor([3, 1])]; + tensor var_601 = reshape(shape = var_600, x = sqrt_s_t_5)[name = tensor("op_601")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_601)[name = tensor("M_5")]; + tensor var_603 = mul(x = qk_5, y = M_5)[name = tensor("op_603")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_590)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_603, y = v_5)[name = tensor("inner_5")]; + tensor var_605_transpose_x_0 = const()[name = tensor("op_605_transpose_x_0"), val = tensor(false)]; + tensor var_605_transpose_y_0 = const()[name = tensor("op_605_transpose_y_0"), val = tensor(false)]; + tensor var_605 = matmul(transpose_x = var_605_transpose_x_0, transpose_y = var_605_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_605")]; + tensor var_606 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_606")]; + tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 1, 3, 1])]; + tensor var_608 = reshape(shape = var_607, x = var_606)[name = tensor("op_608")]; + tensor cross_5 = mul(x = var_605, y = var_608)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_611 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_611")]; + tensor var_613_transpose_x_1 = const()[name = tensor("op_613_transpose_x_1"), val = tensor(true)]; + tensor var_613_transpose_y_1 = const()[name = tensor("op_613_transpose_y_1"), val = tensor(false)]; + tensor var_613 = matmul(transpose_x = var_613_transpose_x_1, transpose_y = var_613_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_613")]; + tensor new_kv_unnorm_5 = add(x = var_611, y = var_613)[name = tensor("new_kv_unnorm_5")]; + tensor var_615 = const()[name = tensor("op_615"), val = tensor(0x1.8p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_615)[name = tensor("new_scale_5")]; + tensor var_617 = sqrt(x = new_scale_5)[name = tensor("op_617")]; + tensor var_618 = real_div(x = new_kv_unnorm_5, y = var_617)[name = tensor("op_618")]; + tensor var_619_perm_0 = const()[name = tensor("op_619_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_619 = transpose(perm = var_619_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_619)[name = tensor("out_15")]; + tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 3, 256])]; + tensor out_17 = reshape(shape = var_623, x = out_15)[name = tensor("out_17")]; + tensor var_625 = silu(x = input_97)[name = tensor("op_625")]; + tensor input_99 = mul(x = var_625, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_17_begin_0 = const()[name = tensor("window_17_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_17_end_0 = const()[name = tensor("window_17_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_17_end_mask_0 = const()[name = tensor("window_17_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_17_squeeze_mask_0 = const()[name = tensor("window_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_17 = slice_by_index(begin = window_17_begin_0, end = window_17_end_0, end_mask = window_17_end_mask_0, squeeze_mask = window_17_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_17")]; + tensor var_633_begin_0 = const()[name = tensor("op_633_begin_0"), val = tensor([0, 0, 0])]; + tensor var_633_end_0 = const()[name = tensor("op_633_end_0"), val = tensor([1, 1, 256])]; + tensor var_633_end_mask_0 = const()[name = tensor("op_633_end_mask_0"), val = tensor([true, false, true])]; + tensor var_633 = slice_by_index(begin = var_633_begin_0, end = var_633_end_0, end_mask = var_633_end_mask_0, x = x_15)[name = tensor("op_633")]; + tensor var_636_begin_0 = const()[name = tensor("op_636_begin_0"), val = tensor([0, 1, 0])]; + tensor var_636_end_0 = const()[name = tensor("op_636_end_0"), val = tensor([1, 16, 256])]; + tensor var_636_end_mask_0 = const()[name = tensor("op_636_end_mask_0"), val = tensor([true, true, true])]; + tensor var_636 = slice_by_index(begin = var_636_begin_0, end = var_636_end_0, end_mask = var_636_end_mask_0, x = window_17)[name = tensor("op_636")]; + tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; + tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_636, var_633))[name = tensor("window_19")]; + tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 1, 0])]; + tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 2, 256])]; + tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, false, true])]; + tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = x_15)[name = tensor("op_641")]; + tensor var_644_begin_0 = const()[name = tensor("op_644_begin_0"), val = tensor([0, 1, 0])]; + tensor var_644_end_0 = const()[name = tensor("op_644_end_0"), val = tensor([1, 16, 256])]; + tensor var_644_end_mask_0 = const()[name = tensor("op_644_end_mask_0"), val = tensor([true, true, true])]; + tensor var_644 = slice_by_index(begin = var_644_begin_0, end = var_644_end_0, end_mask = var_644_end_mask_0, x = window_19)[name = tensor("op_644")]; + tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; + tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_644, var_641))[name = tensor("window_21")]; + tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 2, 0])]; + tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; + tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, true, true])]; + tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; + tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; + tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; + tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; + tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; + tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_19, window_21, window_23))[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_677_split_sizes_0 = const()[name = tensor("op_677_split_sizes_0"), val = tensor([256, 256])]; + tensor var_677_axis_0 = const()[name = tensor("op_677_axis_0"), val = tensor(1)]; + tensor var_677_0, tensor var_677_1 = split(axis = var_677_axis_0, split_sizes = var_677_split_sizes_0, x = inputs_23)[name = tensor("op_677")]; + tensor var_679 = sigmoid(x = var_677_1)[name = tensor("op_679")]; + tensor inputs_25 = mul(x = var_677_0, y = var_679)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([3, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_710_begin_0 = const()[name = tensor("op_710_begin_0"), val = tensor([0, -1, 0])]; + tensor var_710_end_0 = const()[name = tensor("op_710_end_0"), val = tensor([3, 16, 256])]; + tensor var_710_end_mask_0 = const()[name = tensor("op_710_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_710 = slice_by_index(begin = var_710_begin_0, end = var_710_end_0, end_mask = var_710_end_mask_0, x = conv_out_5)[name = tensor("op_710")]; + tensor var_712_perm_0 = const()[name = tensor("op_712_perm_0"), val = tensor([1, 0, 2])]; + tensor var_712 = transpose(perm = var_712_perm_0, x = var_710)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_712)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_735 = const()[name = tensor("op_735"), val = tensor(0x1p-1)]; + tensor var_736 = mul(x = input_119, y = var_735)[name = tensor("op_736")]; + tensor input_121 = add(x = var_736, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor(0x1p-1)]; + tensor var_766 = mul(x = input_131, y = var_765)[name = tensor("op_766")]; + tensor input_133 = add(x = var_766, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_780 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 3, 4, 64])]; + tensor var_782 = reshape(shape = var_781, x = var_780)[name = tensor("op_782")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_786 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_787 = const()[name = tensor("op_787"), val = tensor(0x1p-3)]; + tensor var_788 = mul(x = var_786, y = var_787)[name = tensor("op_788")]; + tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 3, 4, 64])]; + tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_794 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_795 = const()[name = tensor("op_795"), val = tensor([1, 3, 4, 64])]; + tensor var_796 = reshape(shape = var_795, x = var_794)[name = tensor("op_796")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_790)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_782)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_806 = const()[name = tensor("op_806"), val = tensor([3, 1])]; + tensor var_807 = reshape(shape = var_806, x = sqrt_s_t_7)[name = tensor("op_807")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_807)[name = tensor("M_7")]; + tensor var_809 = mul(x = qk_7, y = M_7)[name = tensor("op_809")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_796)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_809, y = v_7)[name = tensor("inner_7")]; + tensor var_811_transpose_x_0 = const()[name = tensor("op_811_transpose_x_0"), val = tensor(false)]; + tensor var_811_transpose_y_0 = const()[name = tensor("op_811_transpose_y_0"), val = tensor(false)]; + tensor var_811 = matmul(transpose_x = var_811_transpose_x_0, transpose_y = var_811_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_811")]; + tensor var_812 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_812")]; + tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 1, 3, 1])]; + tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; + tensor cross_7 = mul(x = var_811, y = var_814)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_817 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_817")]; + tensor var_819_transpose_x_1 = const()[name = tensor("op_819_transpose_x_1"), val = tensor(true)]; + tensor var_819_transpose_y_1 = const()[name = tensor("op_819_transpose_y_1"), val = tensor(false)]; + tensor var_819 = matmul(transpose_x = var_819_transpose_x_1, transpose_y = var_819_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_819")]; + tensor new_kv_unnorm_7 = add(x = var_817, y = var_819)[name = tensor("new_kv_unnorm_7")]; + tensor var_821 = const()[name = tensor("op_821"), val = tensor(0x1.8p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_821)[name = tensor("new_scale_7")]; + tensor var_823 = sqrt(x = new_scale_7)[name = tensor("op_823")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_823)[name = tensor("nkv_1")]; + tensor var_825_perm_0 = const()[name = tensor("op_825_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_825 = transpose(perm = var_825_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_825)[name = tensor("out_21")]; + tensor var_829 = const()[name = tensor("op_829"), val = tensor([1, 3, 256])]; + tensor out_23 = reshape(shape = var_829, x = out_21)[name = tensor("out_23")]; + tensor var_831 = silu(x = input_137)[name = tensor("op_831")]; + tensor input_139 = mul(x = var_831, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; + tensor var_839_begin_0 = const()[name = tensor("op_839_begin_0"), val = tensor([0, 0, 0])]; + tensor var_839_end_0 = const()[name = tensor("op_839_end_0"), val = tensor([1, 1, 256])]; + tensor var_839_end_mask_0 = const()[name = tensor("op_839_end_mask_0"), val = tensor([true, false, true])]; + tensor var_839 = slice_by_index(begin = var_839_begin_0, end = var_839_end_0, end_mask = var_839_end_mask_0, x = x_21)[name = tensor("op_839")]; + tensor var_842_begin_0 = const()[name = tensor("op_842_begin_0"), val = tensor([0, 1, 0])]; + tensor var_842_end_0 = const()[name = tensor("op_842_end_0"), val = tensor([1, 16, 256])]; + tensor var_842_end_mask_0 = const()[name = tensor("op_842_end_mask_0"), val = tensor([true, true, true])]; + tensor var_842 = slice_by_index(begin = var_842_begin_0, end = var_842_end_0, end_mask = var_842_end_mask_0, x = window_25)[name = tensor("op_842")]; + tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; + tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_842, var_839))[name = tensor("window_27")]; + tensor var_847_begin_0 = const()[name = tensor("op_847_begin_0"), val = tensor([0, 1, 0])]; + tensor var_847_end_0 = const()[name = tensor("op_847_end_0"), val = tensor([1, 2, 256])]; + tensor var_847_end_mask_0 = const()[name = tensor("op_847_end_mask_0"), val = tensor([true, false, true])]; + tensor var_847 = slice_by_index(begin = var_847_begin_0, end = var_847_end_0, end_mask = var_847_end_mask_0, x = x_21)[name = tensor("op_847")]; + tensor var_850_begin_0 = const()[name = tensor("op_850_begin_0"), val = tensor([0, 1, 0])]; + tensor var_850_end_0 = const()[name = tensor("op_850_end_0"), val = tensor([1, 16, 256])]; + tensor var_850_end_mask_0 = const()[name = tensor("op_850_end_mask_0"), val = tensor([true, true, true])]; + tensor var_850 = slice_by_index(begin = var_850_begin_0, end = var_850_end_0, end_mask = var_850_end_mask_0, x = window_27)[name = tensor("op_850")]; + tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; + tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_850, var_847))[name = tensor("window_29")]; + tensor var_855_begin_0 = const()[name = tensor("op_855_begin_0"), val = tensor([0, 2, 0])]; + tensor var_855_end_0 = const()[name = tensor("op_855_end_0"), val = tensor([1, 1, 256])]; + tensor var_855_end_mask_0 = const()[name = tensor("op_855_end_mask_0"), val = tensor([true, true, true])]; + tensor var_855 = slice_by_index(begin = var_855_begin_0, end = var_855_end_0, end_mask = var_855_end_mask_0, x = x_21)[name = tensor("op_855")]; + tensor var_858_begin_0 = const()[name = tensor("op_858_begin_0"), val = tensor([0, 1, 0])]; + tensor var_858_end_0 = const()[name = tensor("op_858_end_0"), val = tensor([1, 16, 256])]; + tensor var_858_end_mask_0 = const()[name = tensor("op_858_end_mask_0"), val = tensor([true, true, true])]; + tensor var_858 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = window_29)[name = tensor("op_858")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_858, var_855))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_27, window_29, window))[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_883_split_sizes_0 = const()[name = tensor("op_883_split_sizes_0"), val = tensor([256, 256])]; + tensor var_883_axis_0 = const()[name = tensor("op_883_axis_0"), val = tensor(1)]; + tensor var_883_0, tensor var_883_1 = split(axis = var_883_axis_0, split_sizes = var_883_split_sizes_0, x = inputs_33)[name = tensor("op_883")]; + tensor var_885 = sigmoid(x = var_883_1)[name = tensor("op_885")]; + tensor inputs_35 = mul(x = var_883_0, y = var_885)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([3, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_916_begin_0 = const()[name = tensor("op_916_begin_0"), val = tensor([0, -1, 0])]; + tensor var_916_end_0 = const()[name = tensor("op_916_end_0"), val = tensor([3, 16, 256])]; + tensor var_916_end_mask_0 = const()[name = tensor("op_916_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_916 = slice_by_index(begin = var_916_begin_0, end = var_916_end_0, end_mask = var_916_end_mask_0, x = conv_out_7)[name = tensor("op_916")]; + tensor var_918_perm_0 = const()[name = tensor("op_918_perm_0"), val = tensor([1, 0, 2])]; + tensor var_918 = transpose(perm = var_918_perm_0, x = var_916)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_918)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_941 = const()[name = tensor("op_941"), val = tensor(0x1p-1)]; + tensor var_942 = mul(x = input_159, y = var_941)[name = tensor("op_942")]; + tensor input_161 = add(x = var_942, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_960_begin_0 = const()[name = tensor("op_960_begin_0"), val = tensor([0, 0, 3])]; + tensor var_960_end_0 = const()[name = tensor("op_960_end_0"), val = tensor([1, 256, 21])]; + tensor var_960_end_mask_0 = const()[name = tensor("op_960_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_960_begin_0, end = var_960_end_0, end_mask = var_960_end_mask_0, x = cat)[name = tensor("op_960")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_963 = reduce_l2_norm(axes = var_962, keep_dims = var_29, x = input_163)[name = tensor("op_963")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_963)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_967_axis_0, values = (var_206, var_412, var_618, nkv_1))[name = tensor("op_967")]; + tensor var_969_axis_0 = const()[name = tensor("op_969_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_969_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_969")]; + tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_971_axis_0, values = (window_7, window_15, window_23, window))[name = tensor("op_971")]; + tensor var_980 = const()[name = tensor("op_980"), val = tensor(0x1.5798eep-27)]; + tensor var_985 = const()[name = tensor("op_985"), val = tensor(0x1.4f8b58p-17)]; + tensor var_987 = const()[name = tensor("op_987"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_988 = const()[name = tensor("op_988"), val = tensor(true)]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor(0x1p+0)]; + tensor var_994 = const()[name = tensor("op_994"), val = tensor(-1)]; + tensor var_1000 = const()[name = tensor("op_1000"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_1062_axes_0 = const()[name = tensor("op_1062_axes_0"), val = tensor([2])]; + tensor var_1062 = expand_dims(axes = var_1062_axes_0, x = emb)[name = tensor("op_1062")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 6, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1062)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_994, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1070_perm_0 = const()[name = tensor("op_1070_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1074 = const()[name = tensor("op_1074"), val = tensor([6, 3, 256])]; + tensor var_1070 = transpose(perm = var_1070_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1074, x = var_1070)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 6, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1082 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1083 = const()[name = tensor("op_1083"), val = tensor([6, 3, 4, 64])]; + tensor var_1084 = reshape(shape = var_1083, x = var_1082)[name = tensor("op_1084")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1088 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1089 = const()[name = tensor("op_1089"), val = tensor(0x1p-3)]; + tensor var_1090 = mul(x = var_1088, y = var_1089)[name = tensor("op_1090")]; + tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([6, 3, 4, 64])]; + tensor var_1092 = reshape(shape = var_1091, x = var_1090)[name = tensor("op_1092")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1096 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([6, 3, 4, 64])]; + tensor var_1098 = reshape(shape = var_1097, x = var_1096)[name = tensor("op_1098")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_1000, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_990, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1092)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1084)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([1, 3])]; + tensor var_1111 = reshape(shape = var_1110, x = valid_mask)[name = tensor("op_1111")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1111)[name = tensor("causal_with_valid_1")]; + tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([3, 1])]; + tensor var_1114 = reshape(shape = var_1113, x = sqrt_s_t_9)[name = tensor("op_1114")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1114)[name = tensor("M_9")]; + tensor var_1116 = mul(x = qk_9, y = M_9)[name = tensor("op_1116")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1098)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1116, y = v_9)[name = tensor("inner_9")]; + tensor var_1118_transpose_x_0 = const()[name = tensor("op_1118_transpose_x_0"), val = tensor(false)]; + tensor var_1118_transpose_y_0 = const()[name = tensor("op_1118_transpose_y_0"), val = tensor(false)]; + tensor var_1118 = matmul(transpose_x = var_1118_transpose_x_0, transpose_y = var_1118_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1118")]; + tensor var_1119 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1119")]; + tensor var_1120 = const()[name = tensor("op_1120"), val = tensor([1, 1, 3, 1])]; + tensor var_1121 = reshape(shape = var_1120, x = var_1119)[name = tensor("op_1121")]; + tensor cross_9 = mul(x = var_1118, y = var_1121)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1124 = const()[name = tensor("op_1124"), val = tensor([1, 1, 3, 1])]; + tensor var_1125 = reshape(shape = var_1124, x = valid_mask)[name = tensor("op_1125")]; + tensor v_masked_1 = mul(x = v_9, y = var_1125)[name = tensor("v_masked_1")]; + tensor var_1127 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1127")]; + tensor var_1129_transpose_x_1 = const()[name = tensor("op_1129_transpose_x_1"), val = tensor(true)]; + tensor var_1129_transpose_y_1 = const()[name = tensor("op_1129_transpose_y_1"), val = tensor(false)]; + tensor var_1129 = matmul(transpose_x = var_1129_transpose_x_1, transpose_y = var_1129_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1129")]; + tensor new_kv_unnorm_9 = add(x = var_1127, y = var_1129)[name = tensor("new_kv_unnorm_9")]; + tensor var_1131_keep_dims_0 = const()[name = tensor("op_1131_keep_dims_0"), val = tensor(false)]; + tensor var_1131 = reduce_sum(keep_dims = var_1131_keep_dims_0, x = valid_mask)[name = tensor("op_1131")]; + tensor var_1132 = const()[name = tensor("op_1132"), val = tensor([1])]; + tensor var_1133 = reshape(shape = var_1132, x = var_1131)[name = tensor("op_1133")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1133)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_990, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1137 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1137")]; + tensor var_1138_perm_0 = const()[name = tensor("op_1138_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1138 = transpose(perm = var_1138_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_987, x = var_1138)[name = tensor("out_27")]; + tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([6, 3, 256])]; + tensor out_29 = reshape(shape = var_1142, x = out_27)[name = tensor("out_29")]; + tensor var_1144 = silu(x = input_169)[name = tensor("op_1144")]; + tensor input_171 = mul(x = var_1144, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_985, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 6, 3, 256])]; + tensor var_1155 = reshape(shape = var_1154, x = xt_1)[name = tensor("op_1155")]; + tensor var_1156_perm_0 = const()[name = tensor("op_1156_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1159 = const()[name = tensor("op_1159"), val = tensor([3, 6, 256])]; + tensor var_1156 = transpose(perm = var_1156_perm_0, x = var_1155)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1159, x = var_1156)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1182 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([6, 3, 3, 256])]; + tensor var_1184 = reshape(shape = concat_1, x = var_1182)[name = tensor("op_1184")]; + tensor var_1185_axes_0 = const()[name = tensor("op_1185_axes_0"), val = tensor([0])]; + tensor var_1185 = expand_dims(axes = var_1185_axes_0, x = var_1184)[name = tensor("op_1185")]; + tensor var_1186_perm_0 = const()[name = tensor("op_1186_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1187_axes_0 = const()[name = tensor("op_1187_axes_0"), val = tensor([-2])]; + tensor var_1186 = transpose(perm = var_1186_perm_0, x = var_1185)[name = tensor("transpose_21")]; + tensor var_1187 = squeeze(axes = var_1187_axes_0, x = var_1186)[name = tensor("op_1187")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 6, 3, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1187)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 6, 3, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1187)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 6, 3, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1187)[name = tensor("v_11")]; + tensor var_1195 = const()[name = tensor("op_1195"), val = tensor([6, 12, 64])]; + tensor var_1196 = reshape(shape = var_1195, x = q_11)[name = tensor("op_1196")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1202 = const()[name = tensor("op_1202"), val = tensor([6, 12, 64])]; + tensor var_1203 = reshape(shape = var_1202, x = k_11)[name = tensor("op_1203")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1209 = const()[name = tensor("op_1209"), val = tensor([6, 12, 64])]; + tensor var_1210 = reshape(shape = var_1209, x = v_11)[name = tensor("op_1210")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1213 = const()[name = tensor("op_1213"), val = tensor([3, 4, 6, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1196)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1213, x = q_13)[name = tensor("q_15")]; + tensor var_1215 = const()[name = tensor("op_1215"), val = tensor([3, 4, 6, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1203)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1215, x = k_13)[name = tensor("k_15")]; + tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([3, 4, 6, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1210)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1217, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1220 = const()[name = tensor("op_1220"), val = tensor([2, 0, 1, 3])]; + tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([18, 256])]; + tensor var_1221 = transpose(perm = var_1220, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1225, x = var_1221)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1229 = const()[name = tensor("op_1229"), val = tensor([6, 3, 256])]; + tensor attn_output_7 = reshape(shape = var_1229, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_985, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_985, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([1, 3, 6, 256])]; + tensor x_31 = reshape(shape = var_1249, x = xt_3)[name = tensor("x_31")]; + tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1255 = const()[name = tensor("op_1255"), val = tensor([6, 3, 256])]; + tensor var_1251 = transpose(perm = var_1251_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1255, x = var_1251)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 6, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1263 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1264 = const()[name = tensor("op_1264"), val = tensor([6, 3, 4, 64])]; + tensor var_1265 = reshape(shape = var_1264, x = var_1263)[name = tensor("op_1265")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1269 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1270 = const()[name = tensor("op_1270"), val = tensor(0x1p-3)]; + tensor var_1271 = mul(x = var_1269, y = var_1270)[name = tensor("op_1271")]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([6, 3, 4, 64])]; + tensor var_1273 = reshape(shape = var_1272, x = var_1271)[name = tensor("op_1273")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1277 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([6, 3, 4, 64])]; + tensor var_1279 = reshape(shape = var_1278, x = var_1277)[name = tensor("op_1279")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_990, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1273)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1265)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([3, 1])]; + tensor var_1295 = reshape(shape = var_1294, x = sqrt_s_t)[name = tensor("op_1295")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1295)[name = tensor("M")]; + tensor var_1297 = mul(x = qk, y = M)[name = tensor("op_1297")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1279)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1297, y = v_17)[name = tensor("inner")]; + tensor var_1299_transpose_x_0 = const()[name = tensor("op_1299_transpose_x_0"), val = tensor(false)]; + tensor var_1299_transpose_y_0 = const()[name = tensor("op_1299_transpose_y_0"), val = tensor(false)]; + tensor var_1299 = matmul(transpose_x = var_1299_transpose_x_0, transpose_y = var_1299_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1299")]; + tensor var_1300 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1300")]; + tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([1, 1, 3, 1])]; + tensor var_1302 = reshape(shape = var_1301, x = var_1300)[name = tensor("op_1302")]; + tensor cross = mul(x = var_1299, y = var_1302)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1125)[name = tensor("v_masked")]; + tensor var_1308 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1308")]; + tensor var_1310_transpose_x_1 = const()[name = tensor("op_1310_transpose_x_1"), val = tensor(true)]; + tensor var_1310_transpose_y_1 = const()[name = tensor("op_1310_transpose_y_1"), val = tensor(false)]; + tensor var_1310 = matmul(transpose_x = var_1310_transpose_x_1, transpose_y = var_1310_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1310")]; + tensor new_kv_unnorm = add(x = var_1308, y = var_1310)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1133)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_990, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1319_perm_0 = const()[name = tensor("op_1319_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1319 = transpose(perm = var_1319_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_987, x = var_1319)[name = tensor("out_33")]; + tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([6, 3, 256])]; + tensor out = reshape(shape = var_1323, x = out_33)[name = tensor("out")]; + tensor var_1325 = silu(x = input_187)[name = tensor("op_1325")]; + tensor input_189 = mul(x = var_1325, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_985, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor([1, 6, 3, 256])]; + tensor var_1336 = reshape(shape = var_1335, x = xt_5)[name = tensor("op_1336")]; + tensor var_1337_perm_0 = const()[name = tensor("op_1337_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1340 = const()[name = tensor("op_1340"), val = tensor([3, 6, 256])]; + tensor var_1337 = transpose(perm = var_1337_perm_0, x = var_1336)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1340, x = var_1337)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1363 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([6, 3, 3, 256])]; + tensor var_1365 = reshape(shape = concat_2, x = var_1363)[name = tensor("op_1365")]; + tensor var_1366_axes_0 = const()[name = tensor("op_1366_axes_0"), val = tensor([0])]; + tensor var_1366 = expand_dims(axes = var_1366_axes_0, x = var_1365)[name = tensor("op_1366")]; + tensor var_1367_perm_0 = const()[name = tensor("op_1367_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1368_axes_0 = const()[name = tensor("op_1368_axes_0"), val = tensor([-2])]; + tensor var_1367 = transpose(perm = var_1367_perm_0, x = var_1366)[name = tensor("transpose_8")]; + tensor var_1368 = squeeze(axes = var_1368_axes_0, x = var_1367)[name = tensor("op_1368")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 6, 3, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1368)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 6, 3, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1368)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 6, 3, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1368)[name = tensor("v_19")]; + tensor var_1376 = const()[name = tensor("op_1376"), val = tensor([6, 12, 64])]; + tensor var_1377 = reshape(shape = var_1376, x = q_19)[name = tensor("op_1377")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([6, 12, 64])]; + tensor var_1384 = reshape(shape = var_1383, x = k_19)[name = tensor("op_1384")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([6, 12, 64])]; + tensor var_1391 = reshape(shape = var_1390, x = v_19)[name = tensor("op_1391")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1394 = const()[name = tensor("op_1394"), val = tensor([3, 4, 6, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1377)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1394, x = q_21)[name = tensor("q")]; + tensor var_1396 = const()[name = tensor("op_1396"), val = tensor([3, 4, 6, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1384)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1396, x = k_21)[name = tensor("k")]; + tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([3, 4, 6, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1391)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1398, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1401 = const()[name = tensor("op_1401"), val = tensor([2, 0, 1, 3])]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([18, 256])]; + tensor var_1402 = transpose(perm = var_1401, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1406, x = var_1402)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1410 = const()[name = tensor("op_1410"), val = tensor([6, 3, 256])]; + tensor attn_output = reshape(shape = var_1410, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_985, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_985, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([1, 3, 6, 256])]; + tensor input = reshape(shape = var_1430, x = xt)[name = tensor("input")]; + tensor var_1432 = const()[name = tensor("op_1432"), val = tensor([-1])]; + tensor var_1433 = reduce_l2_norm(axes = var_1432, keep_dims = var_988, x = input)[name = tensor("op_1433")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_980, beta = const_42, x = var_1433)[name = tensor("clip_5")]; + tensor var_1435 = real_div(x = input, y = clip_5)[name = tensor("op_1435")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([3, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([3, 256, 6])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1435)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 3, 6])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 3, 5])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1439")]; + tensor var_1441_axis_0 = const()[name = tensor("op_1441_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1441_axis_0, values = (var_1137, nkv))[name = tensor("op_1441")]; + tensor var_1443_axis_0 = const()[name = tensor("op_1443_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1443_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1443")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/weights/weight.bin b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..e37286e7ad795a75fbea3306827ab92979ce47a2 --- /dev/null +++ b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e7226636660d5bb9d45ee9e4cf2c5a53199cb335cd9fa6c894d0b9a5dd65a95 +size 44413824 diff --git a/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..3ee6a266410b4c8f967985e6dd4010971e326dda --- /dev/null +++ b/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c02e235e84b720c6aaf09d92bbac369c38d556327dd8f5f72579d9e76c86d43 +size 185451 diff --git a/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..e37286e7ad795a75fbea3306827ab92979ce47a2 --- /dev/null +++ b/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e7226636660d5bb9d45ee9e4cf2c5a53199cb335cd9fa6c894d0b9a5dd65a95 +size 44413824 diff --git a/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Manifest.json b/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..c9c2cd6e7172b08333ab9eafcbbc60ac652c7e02 --- /dev/null +++ b/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "68330719-A46B-4379-8B85-C9B4E8FD52AF": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "E31DDEE0-26D5-4CE8-AD79-31130F2DF498": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "68330719-A46B-4379-8B85-C9B4E8FD52AF" +} diff --git a/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/analytics/coremldata.bin b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..991ad5aa06a48e1d0f11baf09e0e8defe19c30ba --- /dev/null +++ b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5f8f8f57d8e374772e2d54b8950c8de8c6a3a8cc10450c3832b9bf5e75ce240 +size 243 diff --git a/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/coremldata.bin b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..2e9e19cf1f89eb088959423ab8f66c6e00895ceb --- /dev/null +++ b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71e5da37ce4f16ebb4f5b2a2ef320e4f17aee294a47d49fa31f021a549ed1af3 +size 1292 diff --git a/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/metadata.json b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..02cd719d1f593b541368c72808de5e7b68e572ad --- /dev/null +++ b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND AMI streaming diarizer (pipeline, T=4, max_speakers=4)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 4 × 4)", + "shortDescription" : "", + "shape" : "[1, 4, 4]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 6 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 6, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 68, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 64, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 22, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 47 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 4 × 345)", + "shortDescription" : "", + "shape" : "[1, 4, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 6 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 6, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4)", + "shortDescription" : "", + "shape" : "[4]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"ami\", \"model_label\": \"AMI\", \"variant\": \"pipeline\", \"chunk_size\": 4, \"step_duration_ms\": 400, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 4, \"max_nspks\": 6, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_ami_400ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/model.mil b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..189cd5c18852bac2bde041df8ebc35de2ca1a491 --- /dev/null +++ b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/model.mil @@ -0,0 +1,1326 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982080)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983168)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336512)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337600)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338688)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339776)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340864)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345024)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393664)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394752)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443392)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444480)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445568)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446656)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708864)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709952)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972160)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973248)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235456)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236544)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498752)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499840)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762048)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763136)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764224)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766336)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290688)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307136)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308224)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309312)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310400)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311488)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312576)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574784)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575872)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576960)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581120)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629760)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630848)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679488)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680576)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681664)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682752)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683840)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688000)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736640)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737728)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786368)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787456)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788544)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789632)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051840)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052928)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315136)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316224)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578432)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579520)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841728)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842816)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105024)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106112)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107200)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109312)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633664)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650112)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651200)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652288)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653376)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654464)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655552)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917760)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918848)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919936)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924096)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972736)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973824)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022464)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023552)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024640)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025728)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026816)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030976)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079616)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080704)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129344)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130432)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131520)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132608)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394816)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395904)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658112)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659200)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921408)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922496)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184704)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185792)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448000)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449088)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450176)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452288)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976640)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993088)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994176)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995264)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996352)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997440)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998528)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260736)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261824)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262912)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267072)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315712)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316800)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365440)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366528)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367616)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368704)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369792)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373952)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422592)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423680)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472320)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473408)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474496)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475584)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737792)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738880)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001088)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002176)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264384)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265472)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527680)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528768)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790976)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792064)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793152)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795264)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319616)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336064)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337152)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338240)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339328)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340416)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341504)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603712)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604800)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605888)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610048)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658688)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659776)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708416)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709504)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710592)))]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710720)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711808)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236160)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237248)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499456)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500544)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762752)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763840)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026048)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027136)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289344)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290432)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552640)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553728)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554816)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555904)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818112)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821248)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607744)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608832)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609920)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618176)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715392)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716480)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813696)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814784)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815872)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816960)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079168)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080256)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342464)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343552)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605760)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606848)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869056)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870144)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132352)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133440)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134528)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135616)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397824)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400960)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187456)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188544)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189632)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197888)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295104)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296192)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393408)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394496)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 4, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 4, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 4, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([4, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 4, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 4, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; + tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; + tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; + tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; + tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; + tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; + tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; + tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; + tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; + tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; + tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; + tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; + tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; + tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 3, 256])]; + tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, false, true])]; + tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; + tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; + tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; + tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; + tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; + tensor var_245_begin_0 = const()[name = tensor("op_245_begin_0"), val = tensor([0, 3, 0])]; + tensor var_245_end_0 = const()[name = tensor("op_245_end_0"), val = tensor([1, 1, 256])]; + tensor var_245_end_mask_0 = const()[name = tensor("op_245_end_mask_0"), val = tensor([true, true, true])]; + tensor var_245 = slice_by_index(begin = var_245_begin_0, end = var_245_end_0, end_mask = var_245_end_mask_0, x = x_3)[name = tensor("op_245")]; + tensor var_248_begin_0 = const()[name = tensor("op_248_begin_0"), val = tensor([0, 1, 0])]; + tensor var_248_end_0 = const()[name = tensor("op_248_end_0"), val = tensor([1, 16, 256])]; + tensor var_248_end_mask_0 = const()[name = tensor("op_248_end_mask_0"), val = tensor([true, true, true])]; + tensor var_248 = slice_by_index(begin = var_248_begin_0, end = var_248_end_0, end_mask = var_248_end_mask_0, x = window_7)[name = tensor("op_248")]; + tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; + tensor window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_248, var_245))[name = tensor("window_9")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9))[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_273_split_sizes_0 = const()[name = tensor("op_273_split_sizes_0"), val = tensor([256, 256])]; + tensor var_273_axis_0 = const()[name = tensor("op_273_axis_0"), val = tensor(1)]; + tensor var_273_0, tensor var_273_1 = split(axis = var_273_axis_0, split_sizes = var_273_split_sizes_0, x = inputs_3)[name = tensor("op_273")]; + tensor var_275 = sigmoid(x = var_273_1)[name = tensor("op_275")]; + tensor inputs_5 = mul(x = var_273_0, y = var_275)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([4, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([0, -1, 0])]; + tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([4, 16, 256])]; + tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_306 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, x = conv_out_1)[name = tensor("op_306")]; + tensor var_308_perm_0 = const()[name = tensor("op_308_perm_0"), val = tensor([1, 0, 2])]; + tensor var_308 = transpose(perm = var_308_perm_0, x = var_306)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_308)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_331 = const()[name = tensor("op_331"), val = tensor(0x1p-1)]; + tensor var_332 = mul(x = input_39, y = var_331)[name = tensor("op_332")]; + tensor input_41 = add(x = var_332, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor(0x1p-1)]; + tensor var_362 = mul(x = input_51, y = var_361)[name = tensor("op_362")]; + tensor input_53 = add(x = var_362, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_376 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 4, 4, 64])]; + tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_382 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_383 = const()[name = tensor("op_383"), val = tensor(0x1p-3)]; + tensor var_384 = mul(x = var_382, y = var_383)[name = tensor("op_384")]; + tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 4, 4, 64])]; + tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_390 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 4, 4, 64])]; + tensor var_392 = reshape(shape = var_391, x = var_390)[name = tensor("op_392")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_386)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_378)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_402 = const()[name = tensor("op_402"), val = tensor([4, 1])]; + tensor var_403 = reshape(shape = var_402, x = sqrt_s_t_3)[name = tensor("op_403")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_403)[name = tensor("M_3")]; + tensor var_405 = mul(x = qk_3, y = M_3)[name = tensor("op_405")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_392)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_405, y = v_3)[name = tensor("inner_3")]; + tensor var_407_transpose_x_0 = const()[name = tensor("op_407_transpose_x_0"), val = tensor(false)]; + tensor var_407_transpose_y_0 = const()[name = tensor("op_407_transpose_y_0"), val = tensor(false)]; + tensor var_407 = matmul(transpose_x = var_407_transpose_x_0, transpose_y = var_407_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_407")]; + tensor var_408 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_408")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1, 4, 1])]; + tensor var_410 = reshape(shape = var_409, x = var_408)[name = tensor("op_410")]; + tensor cross_3 = mul(x = var_407, y = var_410)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_413 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_413")]; + tensor var_415_transpose_x_1 = const()[name = tensor("op_415_transpose_x_1"), val = tensor(true)]; + tensor var_415_transpose_y_1 = const()[name = tensor("op_415_transpose_y_1"), val = tensor(false)]; + tensor var_415 = matmul(transpose_x = var_415_transpose_x_1, transpose_y = var_415_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_415")]; + tensor new_kv_unnorm_3 = add(x = var_413, y = var_415)[name = tensor("new_kv_unnorm_3")]; + tensor var_417 = const()[name = tensor("op_417"), val = tensor(0x1p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_417)[name = tensor("new_scale_3")]; + tensor var_419 = sqrt(x = new_scale_3)[name = tensor("op_419")]; + tensor var_420 = real_div(x = new_kv_unnorm_3, y = var_419)[name = tensor("op_420")]; + tensor var_421_perm_0 = const()[name = tensor("op_421_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_421 = transpose(perm = var_421_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_421)[name = tensor("out_9")]; + tensor var_425 = const()[name = tensor("op_425"), val = tensor([1, 4, 256])]; + tensor out_11 = reshape(shape = var_425, x = out_9)[name = tensor("out_11")]; + tensor var_427 = silu(x = input_57)[name = tensor("op_427")]; + tensor input_59 = mul(x = var_427, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_11_begin_0 = const()[name = tensor("window_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_11_end_0 = const()[name = tensor("window_11_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_11_end_mask_0 = const()[name = tensor("window_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_11_squeeze_mask_0 = const()[name = tensor("window_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_11 = slice_by_index(begin = window_11_begin_0, end = window_11_end_0, end_mask = window_11_end_mask_0, squeeze_mask = window_11_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_11")]; + tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 0, 0])]; + tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 1, 256])]; + tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; + tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; + tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; + tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; + tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; + tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; + tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; + tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 1, 0])]; + tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 2, 256])]; + tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, false, true])]; + tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; + tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; + tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; + tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; + tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; + tensor var_451_begin_0 = const()[name = tensor("op_451_begin_0"), val = tensor([0, 2, 0])]; + tensor var_451_end_0 = const()[name = tensor("op_451_end_0"), val = tensor([1, 3, 256])]; + tensor var_451_end_mask_0 = const()[name = tensor("op_451_end_mask_0"), val = tensor([true, false, true])]; + tensor var_451 = slice_by_index(begin = var_451_begin_0, end = var_451_end_0, end_mask = var_451_end_mask_0, x = x_9)[name = tensor("op_451")]; + tensor var_454_begin_0 = const()[name = tensor("op_454_begin_0"), val = tensor([0, 1, 0])]; + tensor var_454_end_0 = const()[name = tensor("op_454_end_0"), val = tensor([1, 16, 256])]; + tensor var_454_end_mask_0 = const()[name = tensor("op_454_end_mask_0"), val = tensor([true, true, true])]; + tensor var_454 = slice_by_index(begin = var_454_begin_0, end = var_454_end_0, end_mask = var_454_end_mask_0, x = window_15)[name = tensor("op_454")]; + tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; + tensor window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_454, var_451))[name = tensor("window_17")]; + tensor var_459_begin_0 = const()[name = tensor("op_459_begin_0"), val = tensor([0, 3, 0])]; + tensor var_459_end_0 = const()[name = tensor("op_459_end_0"), val = tensor([1, 1, 256])]; + tensor var_459_end_mask_0 = const()[name = tensor("op_459_end_mask_0"), val = tensor([true, true, true])]; + tensor var_459 = slice_by_index(begin = var_459_begin_0, end = var_459_end_0, end_mask = var_459_end_mask_0, x = x_9)[name = tensor("op_459")]; + tensor var_462_begin_0 = const()[name = tensor("op_462_begin_0"), val = tensor([0, 1, 0])]; + tensor var_462_end_0 = const()[name = tensor("op_462_end_0"), val = tensor([1, 16, 256])]; + tensor var_462_end_mask_0 = const()[name = tensor("op_462_end_mask_0"), val = tensor([true, true, true])]; + tensor var_462 = slice_by_index(begin = var_462_begin_0, end = var_462_end_0, end_mask = var_462_end_mask_0, x = window_17)[name = tensor("op_462")]; + tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; + tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_462, var_459))[name = tensor("window_19")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_13, window_15, window_17, window_19))[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_487_split_sizes_0 = const()[name = tensor("op_487_split_sizes_0"), val = tensor([256, 256])]; + tensor var_487_axis_0 = const()[name = tensor("op_487_axis_0"), val = tensor(1)]; + tensor var_487_0, tensor var_487_1 = split(axis = var_487_axis_0, split_sizes = var_487_split_sizes_0, x = inputs_13)[name = tensor("op_487")]; + tensor var_489 = sigmoid(x = var_487_1)[name = tensor("op_489")]; + tensor inputs_15 = mul(x = var_487_0, y = var_489)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([4, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_520_begin_0 = const()[name = tensor("op_520_begin_0"), val = tensor([0, -1, 0])]; + tensor var_520_end_0 = const()[name = tensor("op_520_end_0"), val = tensor([4, 16, 256])]; + tensor var_520_end_mask_0 = const()[name = tensor("op_520_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_520 = slice_by_index(begin = var_520_begin_0, end = var_520_end_0, end_mask = var_520_end_mask_0, x = conv_out_3)[name = tensor("op_520")]; + tensor var_522_perm_0 = const()[name = tensor("op_522_perm_0"), val = tensor([1, 0, 2])]; + tensor var_522 = transpose(perm = var_522_perm_0, x = var_520)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_522)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_545 = const()[name = tensor("op_545"), val = tensor(0x1p-1)]; + tensor var_546 = mul(x = input_79, y = var_545)[name = tensor("op_546")]; + tensor input_81 = add(x = var_546, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-1)]; + tensor var_576 = mul(x = input_91, y = var_575)[name = tensor("op_576")]; + tensor input_93 = add(x = var_576, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_590 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 4, 4, 64])]; + tensor var_592 = reshape(shape = var_591, x = var_590)[name = tensor("op_592")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_596 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_597 = const()[name = tensor("op_597"), val = tensor(0x1p-3)]; + tensor var_598 = mul(x = var_596, y = var_597)[name = tensor("op_598")]; + tensor var_599 = const()[name = tensor("op_599"), val = tensor([1, 4, 4, 64])]; + tensor var_600 = reshape(shape = var_599, x = var_598)[name = tensor("op_600")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_604 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_605 = const()[name = tensor("op_605"), val = tensor([1, 4, 4, 64])]; + tensor var_606 = reshape(shape = var_605, x = var_604)[name = tensor("op_606")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_600)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_592)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_616 = const()[name = tensor("op_616"), val = tensor([4, 1])]; + tensor var_617 = reshape(shape = var_616, x = sqrt_s_t_5)[name = tensor("op_617")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_617)[name = tensor("M_5")]; + tensor var_619 = mul(x = qk_5, y = M_5)[name = tensor("op_619")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_606)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_619, y = v_5)[name = tensor("inner_5")]; + tensor var_621_transpose_x_0 = const()[name = tensor("op_621_transpose_x_0"), val = tensor(false)]; + tensor var_621_transpose_y_0 = const()[name = tensor("op_621_transpose_y_0"), val = tensor(false)]; + tensor var_621 = matmul(transpose_x = var_621_transpose_x_0, transpose_y = var_621_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_621")]; + tensor var_622 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_622")]; + tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 1, 4, 1])]; + tensor var_624 = reshape(shape = var_623, x = var_622)[name = tensor("op_624")]; + tensor cross_5 = mul(x = var_621, y = var_624)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_627 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_627")]; + tensor var_629_transpose_x_1 = const()[name = tensor("op_629_transpose_x_1"), val = tensor(true)]; + tensor var_629_transpose_y_1 = const()[name = tensor("op_629_transpose_y_1"), val = tensor(false)]; + tensor var_629 = matmul(transpose_x = var_629_transpose_x_1, transpose_y = var_629_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_629")]; + tensor new_kv_unnorm_5 = add(x = var_627, y = var_629)[name = tensor("new_kv_unnorm_5")]; + tensor var_631 = const()[name = tensor("op_631"), val = tensor(0x1p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_631)[name = tensor("new_scale_5")]; + tensor var_633 = sqrt(x = new_scale_5)[name = tensor("op_633")]; + tensor var_634 = real_div(x = new_kv_unnorm_5, y = var_633)[name = tensor("op_634")]; + tensor var_635_perm_0 = const()[name = tensor("op_635_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_635 = transpose(perm = var_635_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_635)[name = tensor("out_15")]; + tensor var_639 = const()[name = tensor("op_639"), val = tensor([1, 4, 256])]; + tensor out_17 = reshape(shape = var_639, x = out_15)[name = tensor("out_17")]; + tensor var_641 = silu(x = input_97)[name = tensor("op_641")]; + tensor input_99 = mul(x = var_641, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_21_begin_0 = const()[name = tensor("window_21_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_21_end_0 = const()[name = tensor("window_21_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_21_end_mask_0 = const()[name = tensor("window_21_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_21_squeeze_mask_0 = const()[name = tensor("window_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_21 = slice_by_index(begin = window_21_begin_0, end = window_21_end_0, end_mask = window_21_end_mask_0, squeeze_mask = window_21_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_21")]; + tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 0, 0])]; + tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; + tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, false, true])]; + tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; + tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; + tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; + tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; + tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; + tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; + tensor var_657_begin_0 = const()[name = tensor("op_657_begin_0"), val = tensor([0, 1, 0])]; + tensor var_657_end_0 = const()[name = tensor("op_657_end_0"), val = tensor([1, 2, 256])]; + tensor var_657_end_mask_0 = const()[name = tensor("op_657_end_mask_0"), val = tensor([true, false, true])]; + tensor var_657 = slice_by_index(begin = var_657_begin_0, end = var_657_end_0, end_mask = var_657_end_mask_0, x = x_15)[name = tensor("op_657")]; + tensor var_660_begin_0 = const()[name = tensor("op_660_begin_0"), val = tensor([0, 1, 0])]; + tensor var_660_end_0 = const()[name = tensor("op_660_end_0"), val = tensor([1, 16, 256])]; + tensor var_660_end_mask_0 = const()[name = tensor("op_660_end_mask_0"), val = tensor([true, true, true])]; + tensor var_660 = slice_by_index(begin = var_660_begin_0, end = var_660_end_0, end_mask = var_660_end_mask_0, x = window_23)[name = tensor("op_660")]; + tensor window_25_interleave_0 = const()[name = tensor("window_25_interleave_0"), val = tensor(false)]; + tensor window_25 = concat(axis = var_26, interleave = window_25_interleave_0, values = (var_660, var_657))[name = tensor("window_25")]; + tensor var_665_begin_0 = const()[name = tensor("op_665_begin_0"), val = tensor([0, 2, 0])]; + tensor var_665_end_0 = const()[name = tensor("op_665_end_0"), val = tensor([1, 3, 256])]; + tensor var_665_end_mask_0 = const()[name = tensor("op_665_end_mask_0"), val = tensor([true, false, true])]; + tensor var_665 = slice_by_index(begin = var_665_begin_0, end = var_665_end_0, end_mask = var_665_end_mask_0, x = x_15)[name = tensor("op_665")]; + tensor var_668_begin_0 = const()[name = tensor("op_668_begin_0"), val = tensor([0, 1, 0])]; + tensor var_668_end_0 = const()[name = tensor("op_668_end_0"), val = tensor([1, 16, 256])]; + tensor var_668_end_mask_0 = const()[name = tensor("op_668_end_mask_0"), val = tensor([true, true, true])]; + tensor var_668 = slice_by_index(begin = var_668_begin_0, end = var_668_end_0, end_mask = var_668_end_mask_0, x = window_25)[name = tensor("op_668")]; + tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; + tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_668, var_665))[name = tensor("window_27")]; + tensor var_673_begin_0 = const()[name = tensor("op_673_begin_0"), val = tensor([0, 3, 0])]; + tensor var_673_end_0 = const()[name = tensor("op_673_end_0"), val = tensor([1, 1, 256])]; + tensor var_673_end_mask_0 = const()[name = tensor("op_673_end_mask_0"), val = tensor([true, true, true])]; + tensor var_673 = slice_by_index(begin = var_673_begin_0, end = var_673_end_0, end_mask = var_673_end_mask_0, x = x_15)[name = tensor("op_673")]; + tensor var_676_begin_0 = const()[name = tensor("op_676_begin_0"), val = tensor([0, 1, 0])]; + tensor var_676_end_0 = const()[name = tensor("op_676_end_0"), val = tensor([1, 16, 256])]; + tensor var_676_end_mask_0 = const()[name = tensor("op_676_end_mask_0"), val = tensor([true, true, true])]; + tensor var_676 = slice_by_index(begin = var_676_begin_0, end = var_676_end_0, end_mask = var_676_end_mask_0, x = window_27)[name = tensor("op_676")]; + tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; + tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_676, var_673))[name = tensor("window_29")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_23, window_25, window_27, window_29))[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_701_split_sizes_0 = const()[name = tensor("op_701_split_sizes_0"), val = tensor([256, 256])]; + tensor var_701_axis_0 = const()[name = tensor("op_701_axis_0"), val = tensor(1)]; + tensor var_701_0, tensor var_701_1 = split(axis = var_701_axis_0, split_sizes = var_701_split_sizes_0, x = inputs_23)[name = tensor("op_701")]; + tensor var_703 = sigmoid(x = var_701_1)[name = tensor("op_703")]; + tensor inputs_25 = mul(x = var_701_0, y = var_703)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([4, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_734_begin_0 = const()[name = tensor("op_734_begin_0"), val = tensor([0, -1, 0])]; + tensor var_734_end_0 = const()[name = tensor("op_734_end_0"), val = tensor([4, 16, 256])]; + tensor var_734_end_mask_0 = const()[name = tensor("op_734_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_734 = slice_by_index(begin = var_734_begin_0, end = var_734_end_0, end_mask = var_734_end_mask_0, x = conv_out_5)[name = tensor("op_734")]; + tensor var_736_perm_0 = const()[name = tensor("op_736_perm_0"), val = tensor([1, 0, 2])]; + tensor var_736 = transpose(perm = var_736_perm_0, x = var_734)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_736)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_759 = const()[name = tensor("op_759"), val = tensor(0x1p-1)]; + tensor var_760 = mul(x = input_119, y = var_759)[name = tensor("op_760")]; + tensor input_121 = add(x = var_760, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_789 = const()[name = tensor("op_789"), val = tensor(0x1p-1)]; + tensor var_790 = mul(x = input_131, y = var_789)[name = tensor("op_790")]; + tensor input_133 = add(x = var_790, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_804 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 4, 4, 64])]; + tensor var_806 = reshape(shape = var_805, x = var_804)[name = tensor("op_806")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_810 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_811 = const()[name = tensor("op_811"), val = tensor(0x1p-3)]; + tensor var_812 = mul(x = var_810, y = var_811)[name = tensor("op_812")]; + tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 4, 4, 64])]; + tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_818 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_819 = const()[name = tensor("op_819"), val = tensor([1, 4, 4, 64])]; + tensor var_820 = reshape(shape = var_819, x = var_818)[name = tensor("op_820")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_814)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_806)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_830 = const()[name = tensor("op_830"), val = tensor([4, 1])]; + tensor var_831 = reshape(shape = var_830, x = sqrt_s_t_7)[name = tensor("op_831")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_831)[name = tensor("M_7")]; + tensor var_833 = mul(x = qk_7, y = M_7)[name = tensor("op_833")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_820)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_833, y = v_7)[name = tensor("inner_7")]; + tensor var_835_transpose_x_0 = const()[name = tensor("op_835_transpose_x_0"), val = tensor(false)]; + tensor var_835_transpose_y_0 = const()[name = tensor("op_835_transpose_y_0"), val = tensor(false)]; + tensor var_835 = matmul(transpose_x = var_835_transpose_x_0, transpose_y = var_835_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_835")]; + tensor var_836 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_836")]; + tensor var_837 = const()[name = tensor("op_837"), val = tensor([1, 1, 4, 1])]; + tensor var_838 = reshape(shape = var_837, x = var_836)[name = tensor("op_838")]; + tensor cross_7 = mul(x = var_835, y = var_838)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_841 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_841")]; + tensor var_843_transpose_x_1 = const()[name = tensor("op_843_transpose_x_1"), val = tensor(true)]; + tensor var_843_transpose_y_1 = const()[name = tensor("op_843_transpose_y_1"), val = tensor(false)]; + tensor var_843 = matmul(transpose_x = var_843_transpose_x_1, transpose_y = var_843_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_843")]; + tensor new_kv_unnorm_7 = add(x = var_841, y = var_843)[name = tensor("new_kv_unnorm_7")]; + tensor var_845 = const()[name = tensor("op_845"), val = tensor(0x1p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_845)[name = tensor("new_scale_7")]; + tensor var_847 = sqrt(x = new_scale_7)[name = tensor("op_847")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_847)[name = tensor("nkv_1")]; + tensor var_849_perm_0 = const()[name = tensor("op_849_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_849 = transpose(perm = var_849_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_849)[name = tensor("out_21")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 4, 256])]; + tensor out_23 = reshape(shape = var_853, x = out_21)[name = tensor("out_23")]; + tensor var_855 = silu(x = input_137)[name = tensor("op_855")]; + tensor input_139 = mul(x = var_855, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_31_begin_0 = const()[name = tensor("window_31_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_31_end_0 = const()[name = tensor("window_31_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_31_end_mask_0 = const()[name = tensor("window_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_31_squeeze_mask_0 = const()[name = tensor("window_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_31 = slice_by_index(begin = window_31_begin_0, end = window_31_end_0, end_mask = window_31_end_mask_0, squeeze_mask = window_31_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_31")]; + tensor var_863_begin_0 = const()[name = tensor("op_863_begin_0"), val = tensor([0, 0, 0])]; + tensor var_863_end_0 = const()[name = tensor("op_863_end_0"), val = tensor([1, 1, 256])]; + tensor var_863_end_mask_0 = const()[name = tensor("op_863_end_mask_0"), val = tensor([true, false, true])]; + tensor var_863 = slice_by_index(begin = var_863_begin_0, end = var_863_end_0, end_mask = var_863_end_mask_0, x = x_21)[name = tensor("op_863")]; + tensor var_866_begin_0 = const()[name = tensor("op_866_begin_0"), val = tensor([0, 1, 0])]; + tensor var_866_end_0 = const()[name = tensor("op_866_end_0"), val = tensor([1, 16, 256])]; + tensor var_866_end_mask_0 = const()[name = tensor("op_866_end_mask_0"), val = tensor([true, true, true])]; + tensor var_866 = slice_by_index(begin = var_866_begin_0, end = var_866_end_0, end_mask = var_866_end_mask_0, x = window_31)[name = tensor("op_866")]; + tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; + tensor window_33 = concat(axis = var_26, interleave = window_33_interleave_0, values = (var_866, var_863))[name = tensor("window_33")]; + tensor var_871_begin_0 = const()[name = tensor("op_871_begin_0"), val = tensor([0, 1, 0])]; + tensor var_871_end_0 = const()[name = tensor("op_871_end_0"), val = tensor([1, 2, 256])]; + tensor var_871_end_mask_0 = const()[name = tensor("op_871_end_mask_0"), val = tensor([true, false, true])]; + tensor var_871 = slice_by_index(begin = var_871_begin_0, end = var_871_end_0, end_mask = var_871_end_mask_0, x = x_21)[name = tensor("op_871")]; + tensor var_874_begin_0 = const()[name = tensor("op_874_begin_0"), val = tensor([0, 1, 0])]; + tensor var_874_end_0 = const()[name = tensor("op_874_end_0"), val = tensor([1, 16, 256])]; + tensor var_874_end_mask_0 = const()[name = tensor("op_874_end_mask_0"), val = tensor([true, true, true])]; + tensor var_874 = slice_by_index(begin = var_874_begin_0, end = var_874_end_0, end_mask = var_874_end_mask_0, x = window_33)[name = tensor("op_874")]; + tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; + tensor window_35 = concat(axis = var_26, interleave = window_35_interleave_0, values = (var_874, var_871))[name = tensor("window_35")]; + tensor var_879_begin_0 = const()[name = tensor("op_879_begin_0"), val = tensor([0, 2, 0])]; + tensor var_879_end_0 = const()[name = tensor("op_879_end_0"), val = tensor([1, 3, 256])]; + tensor var_879_end_mask_0 = const()[name = tensor("op_879_end_mask_0"), val = tensor([true, false, true])]; + tensor var_879 = slice_by_index(begin = var_879_begin_0, end = var_879_end_0, end_mask = var_879_end_mask_0, x = x_21)[name = tensor("op_879")]; + tensor var_882_begin_0 = const()[name = tensor("op_882_begin_0"), val = tensor([0, 1, 0])]; + tensor var_882_end_0 = const()[name = tensor("op_882_end_0"), val = tensor([1, 16, 256])]; + tensor var_882_end_mask_0 = const()[name = tensor("op_882_end_mask_0"), val = tensor([true, true, true])]; + tensor var_882 = slice_by_index(begin = var_882_begin_0, end = var_882_end_0, end_mask = var_882_end_mask_0, x = window_35)[name = tensor("op_882")]; + tensor window_37_interleave_0 = const()[name = tensor("window_37_interleave_0"), val = tensor(false)]; + tensor window_37 = concat(axis = var_26, interleave = window_37_interleave_0, values = (var_882, var_879))[name = tensor("window_37")]; + tensor var_887_begin_0 = const()[name = tensor("op_887_begin_0"), val = tensor([0, 3, 0])]; + tensor var_887_end_0 = const()[name = tensor("op_887_end_0"), val = tensor([1, 1, 256])]; + tensor var_887_end_mask_0 = const()[name = tensor("op_887_end_mask_0"), val = tensor([true, true, true])]; + tensor var_887 = slice_by_index(begin = var_887_begin_0, end = var_887_end_0, end_mask = var_887_end_mask_0, x = x_21)[name = tensor("op_887")]; + tensor var_890_begin_0 = const()[name = tensor("op_890_begin_0"), val = tensor([0, 1, 0])]; + tensor var_890_end_0 = const()[name = tensor("op_890_end_0"), val = tensor([1, 16, 256])]; + tensor var_890_end_mask_0 = const()[name = tensor("op_890_end_mask_0"), val = tensor([true, true, true])]; + tensor var_890 = slice_by_index(begin = var_890_begin_0, end = var_890_end_0, end_mask = var_890_end_mask_0, x = window_37)[name = tensor("op_890")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_890, var_887))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_33, window_35, window_37, window))[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_915_split_sizes_0 = const()[name = tensor("op_915_split_sizes_0"), val = tensor([256, 256])]; + tensor var_915_axis_0 = const()[name = tensor("op_915_axis_0"), val = tensor(1)]; + tensor var_915_0, tensor var_915_1 = split(axis = var_915_axis_0, split_sizes = var_915_split_sizes_0, x = inputs_33)[name = tensor("op_915")]; + tensor var_917 = sigmoid(x = var_915_1)[name = tensor("op_917")]; + tensor inputs_35 = mul(x = var_915_0, y = var_917)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([4, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_948_begin_0 = const()[name = tensor("op_948_begin_0"), val = tensor([0, -1, 0])]; + tensor var_948_end_0 = const()[name = tensor("op_948_end_0"), val = tensor([4, 16, 256])]; + tensor var_948_end_mask_0 = const()[name = tensor("op_948_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_948 = slice_by_index(begin = var_948_begin_0, end = var_948_end_0, end_mask = var_948_end_mask_0, x = conv_out_7)[name = tensor("op_948")]; + tensor var_950_perm_0 = const()[name = tensor("op_950_perm_0"), val = tensor([1, 0, 2])]; + tensor var_950 = transpose(perm = var_950_perm_0, x = var_948)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_950)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_973 = const()[name = tensor("op_973"), val = tensor(0x1p-1)]; + tensor var_974 = mul(x = input_159, y = var_973)[name = tensor("op_974")]; + tensor input_161 = add(x = var_974, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_992_begin_0 = const()[name = tensor("op_992_begin_0"), val = tensor([0, 0, 4])]; + tensor var_992_end_0 = const()[name = tensor("op_992_end_0"), val = tensor([1, 256, 22])]; + tensor var_992_end_mask_0 = const()[name = tensor("op_992_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_992_begin_0, end = var_992_end_0, end_mask = var_992_end_mask_0, x = cat)[name = tensor("op_992")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_994 = const()[name = tensor("op_994"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_995 = reduce_l2_norm(axes = var_994, keep_dims = var_29, x = input_163)[name = tensor("op_995")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_995)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_999_axis_0 = const()[name = tensor("op_999_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_999_axis_0, values = (var_206, var_420, var_634, nkv_1))[name = tensor("op_999")]; + tensor var_1001_axis_0 = const()[name = tensor("op_1001_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1001_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1001")]; + tensor var_1003_axis_0 = const()[name = tensor("op_1003_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1003_axis_0, values = (window_9, window_19, window_29, window))[name = tensor("op_1003")]; + tensor var_1012 = const()[name = tensor("op_1012"), val = tensor(0x1.5798eep-27)]; + tensor var_1017 = const()[name = tensor("op_1017"), val = tensor(0x1.4f8b58p-17)]; + tensor var_1019 = const()[name = tensor("op_1019"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_1020 = const()[name = tensor("op_1020"), val = tensor(true)]; + tensor var_1022 = const()[name = tensor("op_1022"), val = tensor(0x1p+0)]; + tensor var_1026 = const()[name = tensor("op_1026"), val = tensor(-1)]; + tensor var_1032 = const()[name = tensor("op_1032"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395584)))]; + tensor var_1094_axes_0 = const()[name = tensor("op_1094_axes_0"), val = tensor([2])]; + tensor var_1094 = expand_dims(axes = var_1094_axes_0, x = emb)[name = tensor("op_1094")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 6, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1094)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_1026, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1102_perm_0 = const()[name = tensor("op_1102_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([6, 4, 256])]; + tensor var_1102 = transpose(perm = var_1102_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1106, x = var_1102)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 6, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1114 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1115 = const()[name = tensor("op_1115"), val = tensor([6, 4, 4, 64])]; + tensor var_1116 = reshape(shape = var_1115, x = var_1114)[name = tensor("op_1116")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1120 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1121 = const()[name = tensor("op_1121"), val = tensor(0x1p-3)]; + tensor var_1122 = mul(x = var_1120, y = var_1121)[name = tensor("op_1122")]; + tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([6, 4, 4, 64])]; + tensor var_1124 = reshape(shape = var_1123, x = var_1122)[name = tensor("op_1124")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1128 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1129 = const()[name = tensor("op_1129"), val = tensor([6, 4, 4, 64])]; + tensor var_1130 = reshape(shape = var_1129, x = var_1128)[name = tensor("op_1130")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_1032, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_1022, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1124)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1116)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([1, 4])]; + tensor var_1143 = reshape(shape = var_1142, x = valid_mask)[name = tensor("op_1143")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1143)[name = tensor("causal_with_valid_1")]; + tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([4, 1])]; + tensor var_1146 = reshape(shape = var_1145, x = sqrt_s_t_9)[name = tensor("op_1146")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1146)[name = tensor("M_9")]; + tensor var_1148 = mul(x = qk_9, y = M_9)[name = tensor("op_1148")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1130)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1148, y = v_9)[name = tensor("inner_9")]; + tensor var_1150_transpose_x_0 = const()[name = tensor("op_1150_transpose_x_0"), val = tensor(false)]; + tensor var_1150_transpose_y_0 = const()[name = tensor("op_1150_transpose_y_0"), val = tensor(false)]; + tensor var_1150 = matmul(transpose_x = var_1150_transpose_x_0, transpose_y = var_1150_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1150")]; + tensor var_1151 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1151")]; + tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 1, 4, 1])]; + tensor var_1153 = reshape(shape = var_1152, x = var_1151)[name = tensor("op_1153")]; + tensor cross_9 = mul(x = var_1150, y = var_1153)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 1, 4, 1])]; + tensor var_1157 = reshape(shape = var_1156, x = valid_mask)[name = tensor("op_1157")]; + tensor v_masked_1 = mul(x = v_9, y = var_1157)[name = tensor("v_masked_1")]; + tensor var_1159 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1159")]; + tensor var_1161_transpose_x_1 = const()[name = tensor("op_1161_transpose_x_1"), val = tensor(true)]; + tensor var_1161_transpose_y_1 = const()[name = tensor("op_1161_transpose_y_1"), val = tensor(false)]; + tensor var_1161 = matmul(transpose_x = var_1161_transpose_x_1, transpose_y = var_1161_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1161")]; + tensor new_kv_unnorm_9 = add(x = var_1159, y = var_1161)[name = tensor("new_kv_unnorm_9")]; + tensor var_1163_keep_dims_0 = const()[name = tensor("op_1163_keep_dims_0"), val = tensor(false)]; + tensor var_1163 = reduce_sum(keep_dims = var_1163_keep_dims_0, x = valid_mask)[name = tensor("op_1163")]; + tensor var_1164 = const()[name = tensor("op_1164"), val = tensor([1])]; + tensor var_1165 = reshape(shape = var_1164, x = var_1163)[name = tensor("op_1165")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1165)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_1022, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1169 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1169")]; + tensor var_1170_perm_0 = const()[name = tensor("op_1170_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1170 = transpose(perm = var_1170_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1019, x = var_1170)[name = tensor("out_27")]; + tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([6, 4, 256])]; + tensor out_29 = reshape(shape = var_1174, x = out_27)[name = tensor("out_29")]; + tensor var_1176 = silu(x = input_169)[name = tensor("op_1176")]; + tensor input_171 = mul(x = var_1176, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1017, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1186 = const()[name = tensor("op_1186"), val = tensor([1, 6, 4, 256])]; + tensor var_1187 = reshape(shape = var_1186, x = xt_1)[name = tensor("op_1187")]; + tensor var_1188_perm_0 = const()[name = tensor("op_1188_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([4, 6, 256])]; + tensor var_1188 = transpose(perm = var_1188_perm_0, x = var_1187)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1191, x = var_1188)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1214 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([6, 4, 3, 256])]; + tensor var_1216 = reshape(shape = concat_1, x = var_1214)[name = tensor("op_1216")]; + tensor var_1217_axes_0 = const()[name = tensor("op_1217_axes_0"), val = tensor([0])]; + tensor var_1217 = expand_dims(axes = var_1217_axes_0, x = var_1216)[name = tensor("op_1217")]; + tensor var_1218_perm_0 = const()[name = tensor("op_1218_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1219_axes_0 = const()[name = tensor("op_1219_axes_0"), val = tensor([-2])]; + tensor var_1218 = transpose(perm = var_1218_perm_0, x = var_1217)[name = tensor("transpose_21")]; + tensor var_1219 = squeeze(axes = var_1219_axes_0, x = var_1218)[name = tensor("op_1219")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 6, 4, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1219)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 6, 4, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1219)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 6, 4, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1219)[name = tensor("v_11")]; + tensor var_1227 = const()[name = tensor("op_1227"), val = tensor([6, 16, 64])]; + tensor var_1228 = reshape(shape = var_1227, x = q_11)[name = tensor("op_1228")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1234 = const()[name = tensor("op_1234"), val = tensor([6, 16, 64])]; + tensor var_1235 = reshape(shape = var_1234, x = k_11)[name = tensor("op_1235")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1241 = const()[name = tensor("op_1241"), val = tensor([6, 16, 64])]; + tensor var_1242 = reshape(shape = var_1241, x = v_11)[name = tensor("op_1242")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([4, 4, 6, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1228)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1245, x = q_13)[name = tensor("q_15")]; + tensor var_1247 = const()[name = tensor("op_1247"), val = tensor([4, 4, 6, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1235)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1247, x = k_13)[name = tensor("k_15")]; + tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([4, 4, 6, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1242)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1249, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([2, 0, 1, 3])]; + tensor var_1257 = const()[name = tensor("op_1257"), val = tensor([24, 256])]; + tensor var_1253 = transpose(perm = var_1252, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1257, x = var_1253)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([6, 4, 256])]; + tensor attn_output_7 = reshape(shape = var_1261, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1017, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1017, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1281 = const()[name = tensor("op_1281"), val = tensor([1, 4, 6, 256])]; + tensor x_31 = reshape(shape = var_1281, x = xt_3)[name = tensor("x_31")]; + tensor var_1283_perm_0 = const()[name = tensor("op_1283_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1287 = const()[name = tensor("op_1287"), val = tensor([6, 4, 256])]; + tensor var_1283 = transpose(perm = var_1283_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1287, x = var_1283)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 6, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1295 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1296 = const()[name = tensor("op_1296"), val = tensor([6, 4, 4, 64])]; + tensor var_1297 = reshape(shape = var_1296, x = var_1295)[name = tensor("op_1297")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1301 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1302 = const()[name = tensor("op_1302"), val = tensor(0x1p-3)]; + tensor var_1303 = mul(x = var_1301, y = var_1302)[name = tensor("op_1303")]; + tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([6, 4, 4, 64])]; + tensor var_1305 = reshape(shape = var_1304, x = var_1303)[name = tensor("op_1305")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1309 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1310 = const()[name = tensor("op_1310"), val = tensor([6, 4, 4, 64])]; + tensor var_1311 = reshape(shape = var_1310, x = var_1309)[name = tensor("op_1311")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_1022, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1305)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1297)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([4, 1])]; + tensor var_1327 = reshape(shape = var_1326, x = sqrt_s_t)[name = tensor("op_1327")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1327)[name = tensor("M")]; + tensor var_1329 = mul(x = qk, y = M)[name = tensor("op_1329")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1311)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1329, y = v_17)[name = tensor("inner")]; + tensor var_1331_transpose_x_0 = const()[name = tensor("op_1331_transpose_x_0"), val = tensor(false)]; + tensor var_1331_transpose_y_0 = const()[name = tensor("op_1331_transpose_y_0"), val = tensor(false)]; + tensor var_1331 = matmul(transpose_x = var_1331_transpose_x_0, transpose_y = var_1331_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1331")]; + tensor var_1332 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1332")]; + tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 1, 4, 1])]; + tensor var_1334 = reshape(shape = var_1333, x = var_1332)[name = tensor("op_1334")]; + tensor cross = mul(x = var_1331, y = var_1334)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1157)[name = tensor("v_masked")]; + tensor var_1340 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1340")]; + tensor var_1342_transpose_x_1 = const()[name = tensor("op_1342_transpose_x_1"), val = tensor(true)]; + tensor var_1342_transpose_y_1 = const()[name = tensor("op_1342_transpose_y_1"), val = tensor(false)]; + tensor var_1342 = matmul(transpose_x = var_1342_transpose_x_1, transpose_y = var_1342_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1342")]; + tensor new_kv_unnorm = add(x = var_1340, y = var_1342)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1165)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_1022, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1351_perm_0 = const()[name = tensor("op_1351_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1351 = transpose(perm = var_1351_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1019, x = var_1351)[name = tensor("out_33")]; + tensor var_1355 = const()[name = tensor("op_1355"), val = tensor([6, 4, 256])]; + tensor out = reshape(shape = var_1355, x = out_33)[name = tensor("out")]; + tensor var_1357 = silu(x = input_187)[name = tensor("op_1357")]; + tensor input_189 = mul(x = var_1357, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1017, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1367 = const()[name = tensor("op_1367"), val = tensor([1, 6, 4, 256])]; + tensor var_1368 = reshape(shape = var_1367, x = xt_5)[name = tensor("op_1368")]; + tensor var_1369_perm_0 = const()[name = tensor("op_1369_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1372 = const()[name = tensor("op_1372"), val = tensor([4, 6, 256])]; + tensor var_1369 = transpose(perm = var_1369_perm_0, x = var_1368)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1372, x = var_1369)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1395 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([6, 4, 3, 256])]; + tensor var_1397 = reshape(shape = concat_2, x = var_1395)[name = tensor("op_1397")]; + tensor var_1398_axes_0 = const()[name = tensor("op_1398_axes_0"), val = tensor([0])]; + tensor var_1398 = expand_dims(axes = var_1398_axes_0, x = var_1397)[name = tensor("op_1398")]; + tensor var_1399_perm_0 = const()[name = tensor("op_1399_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1400_axes_0 = const()[name = tensor("op_1400_axes_0"), val = tensor([-2])]; + tensor var_1399 = transpose(perm = var_1399_perm_0, x = var_1398)[name = tensor("transpose_8")]; + tensor var_1400 = squeeze(axes = var_1400_axes_0, x = var_1399)[name = tensor("op_1400")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 6, 4, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1400)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 6, 4, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1400)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 6, 4, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1400)[name = tensor("v_19")]; + tensor var_1408 = const()[name = tensor("op_1408"), val = tensor([6, 16, 64])]; + tensor var_1409 = reshape(shape = var_1408, x = q_19)[name = tensor("op_1409")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1415 = const()[name = tensor("op_1415"), val = tensor([6, 16, 64])]; + tensor var_1416 = reshape(shape = var_1415, x = k_19)[name = tensor("op_1416")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1422 = const()[name = tensor("op_1422"), val = tensor([6, 16, 64])]; + tensor var_1423 = reshape(shape = var_1422, x = v_19)[name = tensor("op_1423")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1426 = const()[name = tensor("op_1426"), val = tensor([4, 4, 6, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1409)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1426, x = q_21)[name = tensor("q")]; + tensor var_1428 = const()[name = tensor("op_1428"), val = tensor([4, 4, 6, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1416)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1428, x = k_21)[name = tensor("k")]; + tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([4, 4, 6, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1423)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1430, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([2, 0, 1, 3])]; + tensor var_1438 = const()[name = tensor("op_1438"), val = tensor([24, 256])]; + tensor var_1434 = transpose(perm = var_1433, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1438, x = var_1434)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([6, 4, 256])]; + tensor attn_output = reshape(shape = var_1442, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1017, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1017, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1462 = const()[name = tensor("op_1462"), val = tensor([1, 4, 6, 256])]; + tensor input = reshape(shape = var_1462, x = xt)[name = tensor("input")]; + tensor var_1464 = const()[name = tensor("op_1464"), val = tensor([-1])]; + tensor var_1465 = reduce_l2_norm(axes = var_1464, keep_dims = var_1020, x = input)[name = tensor("op_1465")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_1012, beta = const_42, x = var_1465)[name = tensor("clip_5")]; + tensor var_1467 = real_div(x = input, y = clip_5)[name = tensor("op_1467")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([4, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([4, 256, 6])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1467)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 4, 6])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 4, 5])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1471")]; + tensor var_1473_axis_0 = const()[name = tensor("op_1473_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1473_axis_0, values = (var_1169, nkv))[name = tensor("op_1473")]; + tensor var_1475_axis_0 = const()[name = tensor("op_1475_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1475_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1475")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/weights/weight.bin b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..7273476cddb8750b924ed019a5dc6a705d6217f6 --- /dev/null +++ b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37721083b2a8913666e4f80c4f05fa8e8d81a72e136f8a6cf03b4adf1b20c2c4 +size 44420224 diff --git a/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..11bd6cc8bc201ab8dc743d904c5a906849392429 --- /dev/null +++ b/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bf0e98390dbb2eac96d1c7400ce29fa396132e83c940c9f6fb9aa1d779f97c3 +size 191035 diff --git a/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..7273476cddb8750b924ed019a5dc6a705d6217f6 --- /dev/null +++ b/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37721083b2a8913666e4f80c4f05fa8e8d81a72e136f8a6cf03b4adf1b20c2c4 +size 44420224 diff --git a/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Manifest.json b/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..4ea926ed488b5227c9d9a317766f7e6b911916dd --- /dev/null +++ b/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "111C110F-82D7-4E61-A58C-05908C6E61C8": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "A106D900-1931-4022-84A9-F374DF9D3F12": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "A106D900-1931-4022-84A9-F374DF9D3F12" +} diff --git a/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/analytics/coremldata.bin b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..38627aa1b674f9fdb4fc74b0266575b8228d9989 --- /dev/null +++ b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e15806e285f1582d15c27637594c49104f55b3b7b72abdaad4b32313223bee5b +size 243 diff --git a/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/coremldata.bin b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..93ccd031fa1a60d8bc1e9e1900eebe4ed013ded1 --- /dev/null +++ b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77dd81a1ebcc4c55be5fef0d9f8639561671cb3d7b23bed333b23b53fe9a5ab7 +size 1292 diff --git a/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/metadata.json b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..410cbbe3ea4581edc9b16eeffe96a1435ac493c2 --- /dev/null +++ b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND AMI streaming diarizer (pipeline, T=5, max_speakers=4)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 5 × 4)", + "shortDescription" : "", + "shape" : "[1, 5, 4]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 6 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 6, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 68, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 72, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 26, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 47 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 5 × 345)", + "shortDescription" : "", + "shape" : "[1, 5, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 6 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 6, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 5)", + "shortDescription" : "", + "shape" : "[5]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"ami\", \"model_label\": \"AMI\", \"variant\": \"pipeline\", \"chunk_size\": 5, \"step_duration_ms\": 500, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 4, \"max_nspks\": 6, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_ami_500ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/model.mil b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..f062932fd3c10f5f9a6925badb848bf8e06e6c01 --- /dev/null +++ b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/model.mil @@ -0,0 +1,1366 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2, 0x1.4p+2])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982144)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983232)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336576)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337664)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338752)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339840)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340928)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345088)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393728)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394816)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443456)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444544)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445632)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446720)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708928)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7710016)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972224)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973312)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235520)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236608)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498816)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499904)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762112)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763200)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764288)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766400)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290752)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307200)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308288)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309376)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310464)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311552)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312640)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574848)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575936)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9577024)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581184)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629824)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630912)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679552)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680640)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681728)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682816)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683904)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688064)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736704)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737792)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786432)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787520)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788608)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789696)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051904)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052992)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315200)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316288)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578496)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579584)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841792)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842880)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105088)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106176)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107264)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109376)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633728)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650176)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651264)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652352)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653440)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654528)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655616)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917824)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918912)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15920000)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924160)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972800)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973888)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022528)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023616)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024704)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025792)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026880)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18031040)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079680)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080768)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129408)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130496)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131584)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132672)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394880)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395968)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658176)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659264)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921472)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922560)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184768)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185856)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448064)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449152)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450240)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452352)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976704)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993152)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994240)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995328)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996416)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997504)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998592)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260800)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261888)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262976)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267136)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315776)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316864)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365504)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366592)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367680)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368768)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369856)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24374016)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422656)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423744)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472384)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473472)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474560)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475648)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737856)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738944)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001152)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002240)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264448)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265536)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527744)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528832)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791040)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792128)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793216)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795328)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319680)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336128)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337216)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338304)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339392)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340480)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341568)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603776)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604864)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605952)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610112)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658752)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659840)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708480)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709568)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710656)))]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710848)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711936)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236288)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237376)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499584)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500672)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762880)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763968)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026176)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027264)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289472)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290560)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552768)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553856)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554944)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32556032)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818240)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821376)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607872)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608960)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33610048)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618304)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715520)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716608)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813824)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814912)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816000)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37817088)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079296)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080384)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342592)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343680)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605888)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606976)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869184)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870272)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132480)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133568)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134656)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135744)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397952)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39401088)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187584)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188672)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189760)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40198016)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295232)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296320)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393536)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394624)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_18 = const()[name = tensor("op_18"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_21 = const()[name = tensor("op_21"), val = tensor(2)]; + tensor var_24 = const()[name = tensor("op_24"), val = tensor(0)]; + tensor var_27 = const()[name = tensor("op_27"), val = tensor(1)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(0x1.4f8b58p-17)]; + tensor var_30 = const()[name = tensor("op_30"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_29, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_148 = const()[name = tensor("op_148"), val = tensor(0x1p-1)]; + tensor var_149 = mul(x = input_11, y = var_148)[name = tensor("op_149")]; + tensor input_13 = add(x = var_149, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_29, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_163 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_164 = const()[name = tensor("op_164"), val = tensor([1, 5, 4, 64])]; + tensor var_165 = reshape(shape = var_164, x = var_163)[name = tensor("op_165")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_169 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_170 = const()[name = tensor("op_170"), val = tensor(0x1p-3)]; + tensor var_171 = mul(x = var_169, y = var_170)[name = tensor("op_171")]; + tensor var_172 = const()[name = tensor("op_172"), val = tensor([1, 5, 4, 64])]; + tensor var_173 = reshape(shape = var_172, x = var_171)[name = tensor("op_173")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_177 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_178 = const()[name = tensor("op_178"), val = tensor([1, 5, 4, 64])]; + tensor var_179 = reshape(shape = var_178, x = var_177)[name = tensor("op_179")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_173)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_165)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_189 = const()[name = tensor("op_189"), val = tensor([5, 1])]; + tensor var_190 = reshape(shape = var_189, x = sqrt_s_t_1)[name = tensor("op_190")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_190)[name = tensor("M_1")]; + tensor var_192 = mul(x = qk_1, y = M_1)[name = tensor("op_192")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_179)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_192, y = v_1)[name = tensor("inner_1")]; + tensor var_194_transpose_x_0 = const()[name = tensor("op_194_transpose_x_0"), val = tensor(false)]; + tensor var_194_transpose_y_0 = const()[name = tensor("op_194_transpose_y_0"), val = tensor(false)]; + tensor var_194 = matmul(transpose_x = var_194_transpose_x_0, transpose_y = var_194_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_194")]; + tensor var_195 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_195")]; + tensor var_196 = const()[name = tensor("op_196"), val = tensor([1, 1, 5, 1])]; + tensor var_197 = reshape(shape = var_196, x = var_195)[name = tensor("op_197")]; + tensor cross_1 = mul(x = var_194, y = var_197)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_200 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_200")]; + tensor var_202_transpose_x_1 = const()[name = tensor("op_202_transpose_x_1"), val = tensor(true)]; + tensor var_202_transpose_y_1 = const()[name = tensor("op_202_transpose_y_1"), val = tensor(false)]; + tensor var_202 = matmul(transpose_x = var_202_transpose_x_1, transpose_y = var_202_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_202")]; + tensor new_kv_unnorm_1 = add(x = var_200, y = var_202)[name = tensor("new_kv_unnorm_1")]; + tensor var_204 = const()[name = tensor("op_204"), val = tensor(0x1.4p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_204)[name = tensor("new_scale_1")]; + tensor var_206 = sqrt(x = new_scale_1)[name = tensor("op_206")]; + tensor var_207 = real_div(x = new_kv_unnorm_1, y = var_206)[name = tensor("op_207")]; + tensor var_208_perm_0 = const()[name = tensor("op_208_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_208 = transpose(perm = var_208_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_18, x = var_208)[name = tensor("out_3")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor([1, 5, 256])]; + tensor out_5 = reshape(shape = var_212, x = out_3)[name = tensor("out_5")]; + tensor var_214 = silu(x = input_17)[name = tensor("op_214")]; + tensor input_19 = mul(x = var_214, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_222_begin_0 = const()[name = tensor("op_222_begin_0"), val = tensor([0, 0, 0])]; + tensor var_222_end_0 = const()[name = tensor("op_222_end_0"), val = tensor([1, 1, 256])]; + tensor var_222_end_mask_0 = const()[name = tensor("op_222_end_mask_0"), val = tensor([true, false, true])]; + tensor var_222 = slice_by_index(begin = var_222_begin_0, end = var_222_end_0, end_mask = var_222_end_mask_0, x = x_3)[name = tensor("op_222")]; + tensor var_225_begin_0 = const()[name = tensor("op_225_begin_0"), val = tensor([0, 1, 0])]; + tensor var_225_end_0 = const()[name = tensor("op_225_end_0"), val = tensor([1, 16, 256])]; + tensor var_225_end_mask_0 = const()[name = tensor("op_225_end_mask_0"), val = tensor([true, true, true])]; + tensor var_225 = slice_by_index(begin = var_225_begin_0, end = var_225_end_0, end_mask = var_225_end_mask_0, x = window_1)[name = tensor("op_225")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_27, interleave = window_3_interleave_0, values = (var_225, var_222))[name = tensor("window_3")]; + tensor var_230_begin_0 = const()[name = tensor("op_230_begin_0"), val = tensor([0, 1, 0])]; + tensor var_230_end_0 = const()[name = tensor("op_230_end_0"), val = tensor([1, 2, 256])]; + tensor var_230_end_mask_0 = const()[name = tensor("op_230_end_mask_0"), val = tensor([true, false, true])]; + tensor var_230 = slice_by_index(begin = var_230_begin_0, end = var_230_end_0, end_mask = var_230_end_mask_0, x = x_3)[name = tensor("op_230")]; + tensor var_233_begin_0 = const()[name = tensor("op_233_begin_0"), val = tensor([0, 1, 0])]; + tensor var_233_end_0 = const()[name = tensor("op_233_end_0"), val = tensor([1, 16, 256])]; + tensor var_233_end_mask_0 = const()[name = tensor("op_233_end_mask_0"), val = tensor([true, true, true])]; + tensor var_233 = slice_by_index(begin = var_233_begin_0, end = var_233_end_0, end_mask = var_233_end_mask_0, x = window_3)[name = tensor("op_233")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_27, interleave = window_5_interleave_0, values = (var_233, var_230))[name = tensor("window_5")]; + tensor var_238_begin_0 = const()[name = tensor("op_238_begin_0"), val = tensor([0, 2, 0])]; + tensor var_238_end_0 = const()[name = tensor("op_238_end_0"), val = tensor([1, 3, 256])]; + tensor var_238_end_mask_0 = const()[name = tensor("op_238_end_mask_0"), val = tensor([true, false, true])]; + tensor var_238 = slice_by_index(begin = var_238_begin_0, end = var_238_end_0, end_mask = var_238_end_mask_0, x = x_3)[name = tensor("op_238")]; + tensor var_241_begin_0 = const()[name = tensor("op_241_begin_0"), val = tensor([0, 1, 0])]; + tensor var_241_end_0 = const()[name = tensor("op_241_end_0"), val = tensor([1, 16, 256])]; + tensor var_241_end_mask_0 = const()[name = tensor("op_241_end_mask_0"), val = tensor([true, true, true])]; + tensor var_241 = slice_by_index(begin = var_241_begin_0, end = var_241_end_0, end_mask = var_241_end_mask_0, x = window_5)[name = tensor("op_241")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_27, interleave = window_7_interleave_0, values = (var_241, var_238))[name = tensor("window_7")]; + tensor var_246_begin_0 = const()[name = tensor("op_246_begin_0"), val = tensor([0, 3, 0])]; + tensor var_246_end_0 = const()[name = tensor("op_246_end_0"), val = tensor([1, 4, 256])]; + tensor var_246_end_mask_0 = const()[name = tensor("op_246_end_mask_0"), val = tensor([true, false, true])]; + tensor var_246 = slice_by_index(begin = var_246_begin_0, end = var_246_end_0, end_mask = var_246_end_mask_0, x = x_3)[name = tensor("op_246")]; + tensor var_249_begin_0 = const()[name = tensor("op_249_begin_0"), val = tensor([0, 1, 0])]; + tensor var_249_end_0 = const()[name = tensor("op_249_end_0"), val = tensor([1, 16, 256])]; + tensor var_249_end_mask_0 = const()[name = tensor("op_249_end_mask_0"), val = tensor([true, true, true])]; + tensor var_249 = slice_by_index(begin = var_249_begin_0, end = var_249_end_0, end_mask = var_249_end_mask_0, x = window_7)[name = tensor("op_249")]; + tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; + tensor window_9 = concat(axis = var_27, interleave = window_9_interleave_0, values = (var_249, var_246))[name = tensor("window_9")]; + tensor var_254_begin_0 = const()[name = tensor("op_254_begin_0"), val = tensor([0, 4, 0])]; + tensor var_254_end_0 = const()[name = tensor("op_254_end_0"), val = tensor([1, 1, 256])]; + tensor var_254_end_mask_0 = const()[name = tensor("op_254_end_mask_0"), val = tensor([true, true, true])]; + tensor var_254 = slice_by_index(begin = var_254_begin_0, end = var_254_end_0, end_mask = var_254_end_mask_0, x = x_3)[name = tensor("op_254")]; + tensor var_257_begin_0 = const()[name = tensor("op_257_begin_0"), val = tensor([0, 1, 0])]; + tensor var_257_end_0 = const()[name = tensor("op_257_end_0"), val = tensor([1, 16, 256])]; + tensor var_257_end_mask_0 = const()[name = tensor("op_257_end_mask_0"), val = tensor([true, true, true])]; + tensor var_257 = slice_by_index(begin = var_257_begin_0, end = var_257_end_0, end_mask = var_257_end_mask_0, x = window_9)[name = tensor("op_257")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_27, interleave = window_11_interleave_0, values = (var_257, var_254))[name = tensor("window_11")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_24, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9, window_11))[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_282_split_sizes_0 = const()[name = tensor("op_282_split_sizes_0"), val = tensor([256, 256])]; + tensor var_282_axis_0 = const()[name = tensor("op_282_axis_0"), val = tensor(1)]; + tensor var_282_0, tensor var_282_1 = split(axis = var_282_axis_0, split_sizes = var_282_split_sizes_0, x = inputs_3)[name = tensor("op_282")]; + tensor var_284 = sigmoid(x = var_282_1)[name = tensor("op_284")]; + tensor inputs_5 = mul(x = var_282_0, y = var_284)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([5, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_315_begin_0 = const()[name = tensor("op_315_begin_0"), val = tensor([0, -1, 0])]; + tensor var_315_end_0 = const()[name = tensor("op_315_end_0"), val = tensor([5, 16, 256])]; + tensor var_315_end_mask_0 = const()[name = tensor("op_315_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_315 = slice_by_index(begin = var_315_begin_0, end = var_315_end_0, end_mask = var_315_end_mask_0, x = conv_out_1)[name = tensor("op_315")]; + tensor var_317_perm_0 = const()[name = tensor("op_317_perm_0"), val = tensor([1, 0, 2])]; + tensor var_317 = transpose(perm = var_317_perm_0, x = var_315)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_317)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor(0x1p-1)]; + tensor var_341 = mul(x = input_39, y = var_340)[name = tensor("op_341")]; + tensor input_41 = add(x = var_341, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_29, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_370 = const()[name = tensor("op_370"), val = tensor(0x1p-1)]; + tensor var_371 = mul(x = input_51, y = var_370)[name = tensor("op_371")]; + tensor input_53 = add(x = var_371, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_29, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_385 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([1, 5, 4, 64])]; + tensor var_387 = reshape(shape = var_386, x = var_385)[name = tensor("op_387")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_391 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_392 = const()[name = tensor("op_392"), val = tensor(0x1p-3)]; + tensor var_393 = mul(x = var_391, y = var_392)[name = tensor("op_393")]; + tensor var_394 = const()[name = tensor("op_394"), val = tensor([1, 5, 4, 64])]; + tensor var_395 = reshape(shape = var_394, x = var_393)[name = tensor("op_395")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_399 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_400 = const()[name = tensor("op_400"), val = tensor([1, 5, 4, 64])]; + tensor var_401 = reshape(shape = var_400, x = var_399)[name = tensor("op_401")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_395)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_387)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([5, 1])]; + tensor var_412 = reshape(shape = var_411, x = sqrt_s_t_3)[name = tensor("op_412")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_412)[name = tensor("M_3")]; + tensor var_414 = mul(x = qk_3, y = M_3)[name = tensor("op_414")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_401)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_414, y = v_3)[name = tensor("inner_3")]; + tensor var_416_transpose_x_0 = const()[name = tensor("op_416_transpose_x_0"), val = tensor(false)]; + tensor var_416_transpose_y_0 = const()[name = tensor("op_416_transpose_y_0"), val = tensor(false)]; + tensor var_416 = matmul(transpose_x = var_416_transpose_x_0, transpose_y = var_416_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_416")]; + tensor var_417 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_417")]; + tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1, 5, 1])]; + tensor var_419 = reshape(shape = var_418, x = var_417)[name = tensor("op_419")]; + tensor cross_3 = mul(x = var_416, y = var_419)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_422 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_422")]; + tensor var_424_transpose_x_1 = const()[name = tensor("op_424_transpose_x_1"), val = tensor(true)]; + tensor var_424_transpose_y_1 = const()[name = tensor("op_424_transpose_y_1"), val = tensor(false)]; + tensor var_424 = matmul(transpose_x = var_424_transpose_x_1, transpose_y = var_424_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_424")]; + tensor new_kv_unnorm_3 = add(x = var_422, y = var_424)[name = tensor("new_kv_unnorm_3")]; + tensor var_426 = const()[name = tensor("op_426"), val = tensor(0x1.4p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_426)[name = tensor("new_scale_3")]; + tensor var_428 = sqrt(x = new_scale_3)[name = tensor("op_428")]; + tensor var_429 = real_div(x = new_kv_unnorm_3, y = var_428)[name = tensor("op_429")]; + tensor var_430_perm_0 = const()[name = tensor("op_430_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_430 = transpose(perm = var_430_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_18, x = var_430)[name = tensor("out_9")]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 5, 256])]; + tensor out_11 = reshape(shape = var_434, x = out_9)[name = tensor("out_11")]; + tensor var_436 = silu(x = input_57)[name = tensor("op_436")]; + tensor input_59 = mul(x = var_436, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; + tensor var_444_begin_0 = const()[name = tensor("op_444_begin_0"), val = tensor([0, 0, 0])]; + tensor var_444_end_0 = const()[name = tensor("op_444_end_0"), val = tensor([1, 1, 256])]; + tensor var_444_end_mask_0 = const()[name = tensor("op_444_end_mask_0"), val = tensor([true, false, true])]; + tensor var_444 = slice_by_index(begin = var_444_begin_0, end = var_444_end_0, end_mask = var_444_end_mask_0, x = x_9)[name = tensor("op_444")]; + tensor var_447_begin_0 = const()[name = tensor("op_447_begin_0"), val = tensor([0, 1, 0])]; + tensor var_447_end_0 = const()[name = tensor("op_447_end_0"), val = tensor([1, 16, 256])]; + tensor var_447_end_mask_0 = const()[name = tensor("op_447_end_mask_0"), val = tensor([true, true, true])]; + tensor var_447 = slice_by_index(begin = var_447_begin_0, end = var_447_end_0, end_mask = var_447_end_mask_0, x = window_13)[name = tensor("op_447")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_27, interleave = window_15_interleave_0, values = (var_447, var_444))[name = tensor("window_15")]; + tensor var_452_begin_0 = const()[name = tensor("op_452_begin_0"), val = tensor([0, 1, 0])]; + tensor var_452_end_0 = const()[name = tensor("op_452_end_0"), val = tensor([1, 2, 256])]; + tensor var_452_end_mask_0 = const()[name = tensor("op_452_end_mask_0"), val = tensor([true, false, true])]; + tensor var_452 = slice_by_index(begin = var_452_begin_0, end = var_452_end_0, end_mask = var_452_end_mask_0, x = x_9)[name = tensor("op_452")]; + tensor var_455_begin_0 = const()[name = tensor("op_455_begin_0"), val = tensor([0, 1, 0])]; + tensor var_455_end_0 = const()[name = tensor("op_455_end_0"), val = tensor([1, 16, 256])]; + tensor var_455_end_mask_0 = const()[name = tensor("op_455_end_mask_0"), val = tensor([true, true, true])]; + tensor var_455 = slice_by_index(begin = var_455_begin_0, end = var_455_end_0, end_mask = var_455_end_mask_0, x = window_15)[name = tensor("op_455")]; + tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; + tensor window_17 = concat(axis = var_27, interleave = window_17_interleave_0, values = (var_455, var_452))[name = tensor("window_17")]; + tensor var_460_begin_0 = const()[name = tensor("op_460_begin_0"), val = tensor([0, 2, 0])]; + tensor var_460_end_0 = const()[name = tensor("op_460_end_0"), val = tensor([1, 3, 256])]; + tensor var_460_end_mask_0 = const()[name = tensor("op_460_end_mask_0"), val = tensor([true, false, true])]; + tensor var_460 = slice_by_index(begin = var_460_begin_0, end = var_460_end_0, end_mask = var_460_end_mask_0, x = x_9)[name = tensor("op_460")]; + tensor var_463_begin_0 = const()[name = tensor("op_463_begin_0"), val = tensor([0, 1, 0])]; + tensor var_463_end_0 = const()[name = tensor("op_463_end_0"), val = tensor([1, 16, 256])]; + tensor var_463_end_mask_0 = const()[name = tensor("op_463_end_mask_0"), val = tensor([true, true, true])]; + tensor var_463 = slice_by_index(begin = var_463_begin_0, end = var_463_end_0, end_mask = var_463_end_mask_0, x = window_17)[name = tensor("op_463")]; + tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; + tensor window_19 = concat(axis = var_27, interleave = window_19_interleave_0, values = (var_463, var_460))[name = tensor("window_19")]; + tensor var_468_begin_0 = const()[name = tensor("op_468_begin_0"), val = tensor([0, 3, 0])]; + tensor var_468_end_0 = const()[name = tensor("op_468_end_0"), val = tensor([1, 4, 256])]; + tensor var_468_end_mask_0 = const()[name = tensor("op_468_end_mask_0"), val = tensor([true, false, true])]; + tensor var_468 = slice_by_index(begin = var_468_begin_0, end = var_468_end_0, end_mask = var_468_end_mask_0, x = x_9)[name = tensor("op_468")]; + tensor var_471_begin_0 = const()[name = tensor("op_471_begin_0"), val = tensor([0, 1, 0])]; + tensor var_471_end_0 = const()[name = tensor("op_471_end_0"), val = tensor([1, 16, 256])]; + tensor var_471_end_mask_0 = const()[name = tensor("op_471_end_mask_0"), val = tensor([true, true, true])]; + tensor var_471 = slice_by_index(begin = var_471_begin_0, end = var_471_end_0, end_mask = var_471_end_mask_0, x = window_19)[name = tensor("op_471")]; + tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; + tensor window_21 = concat(axis = var_27, interleave = window_21_interleave_0, values = (var_471, var_468))[name = tensor("window_21")]; + tensor var_476_begin_0 = const()[name = tensor("op_476_begin_0"), val = tensor([0, 4, 0])]; + tensor var_476_end_0 = const()[name = tensor("op_476_end_0"), val = tensor([1, 1, 256])]; + tensor var_476_end_mask_0 = const()[name = tensor("op_476_end_mask_0"), val = tensor([true, true, true])]; + tensor var_476 = slice_by_index(begin = var_476_begin_0, end = var_476_end_0, end_mask = var_476_end_mask_0, x = x_9)[name = tensor("op_476")]; + tensor var_479_begin_0 = const()[name = tensor("op_479_begin_0"), val = tensor([0, 1, 0])]; + tensor var_479_end_0 = const()[name = tensor("op_479_end_0"), val = tensor([1, 16, 256])]; + tensor var_479_end_mask_0 = const()[name = tensor("op_479_end_mask_0"), val = tensor([true, true, true])]; + tensor var_479 = slice_by_index(begin = var_479_begin_0, end = var_479_end_0, end_mask = var_479_end_mask_0, x = window_21)[name = tensor("op_479")]; + tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; + tensor window_23 = concat(axis = var_27, interleave = window_23_interleave_0, values = (var_479, var_476))[name = tensor("window_23")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_24, interleave = input_61_interleave_0, values = (window_15, window_17, window_19, window_21, window_23))[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_504_split_sizes_0 = const()[name = tensor("op_504_split_sizes_0"), val = tensor([256, 256])]; + tensor var_504_axis_0 = const()[name = tensor("op_504_axis_0"), val = tensor(1)]; + tensor var_504_0, tensor var_504_1 = split(axis = var_504_axis_0, split_sizes = var_504_split_sizes_0, x = inputs_13)[name = tensor("op_504")]; + tensor var_506 = sigmoid(x = var_504_1)[name = tensor("op_506")]; + tensor inputs_15 = mul(x = var_504_0, y = var_506)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([5, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_537_begin_0 = const()[name = tensor("op_537_begin_0"), val = tensor([0, -1, 0])]; + tensor var_537_end_0 = const()[name = tensor("op_537_end_0"), val = tensor([5, 16, 256])]; + tensor var_537_end_mask_0 = const()[name = tensor("op_537_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_537 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = conv_out_3)[name = tensor("op_537")]; + tensor var_539_perm_0 = const()[name = tensor("op_539_perm_0"), val = tensor([1, 0, 2])]; + tensor var_539 = transpose(perm = var_539_perm_0, x = var_537)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_539)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_562 = const()[name = tensor("op_562"), val = tensor(0x1p-1)]; + tensor var_563 = mul(x = input_79, y = var_562)[name = tensor("op_563")]; + tensor input_81 = add(x = var_563, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_29, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_592 = const()[name = tensor("op_592"), val = tensor(0x1p-1)]; + tensor var_593 = mul(x = input_91, y = var_592)[name = tensor("op_593")]; + tensor input_93 = add(x = var_593, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_29, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_607 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_608 = const()[name = tensor("op_608"), val = tensor([1, 5, 4, 64])]; + tensor var_609 = reshape(shape = var_608, x = var_607)[name = tensor("op_609")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_613 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_614 = const()[name = tensor("op_614"), val = tensor(0x1p-3)]; + tensor var_615 = mul(x = var_613, y = var_614)[name = tensor("op_615")]; + tensor var_616 = const()[name = tensor("op_616"), val = tensor([1, 5, 4, 64])]; + tensor var_617 = reshape(shape = var_616, x = var_615)[name = tensor("op_617")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_621 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_622 = const()[name = tensor("op_622"), val = tensor([1, 5, 4, 64])]; + tensor var_623 = reshape(shape = var_622, x = var_621)[name = tensor("op_623")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_617)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_609)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_633 = const()[name = tensor("op_633"), val = tensor([5, 1])]; + tensor var_634 = reshape(shape = var_633, x = sqrt_s_t_5)[name = tensor("op_634")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_634)[name = tensor("M_5")]; + tensor var_636 = mul(x = qk_5, y = M_5)[name = tensor("op_636")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_623)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_636, y = v_5)[name = tensor("inner_5")]; + tensor var_638_transpose_x_0 = const()[name = tensor("op_638_transpose_x_0"), val = tensor(false)]; + tensor var_638_transpose_y_0 = const()[name = tensor("op_638_transpose_y_0"), val = tensor(false)]; + tensor var_638 = matmul(transpose_x = var_638_transpose_x_0, transpose_y = var_638_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_638")]; + tensor var_639 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_639")]; + tensor var_640 = const()[name = tensor("op_640"), val = tensor([1, 1, 5, 1])]; + tensor var_641 = reshape(shape = var_640, x = var_639)[name = tensor("op_641")]; + tensor cross_5 = mul(x = var_638, y = var_641)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_644 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_644")]; + tensor var_646_transpose_x_1 = const()[name = tensor("op_646_transpose_x_1"), val = tensor(true)]; + tensor var_646_transpose_y_1 = const()[name = tensor("op_646_transpose_y_1"), val = tensor(false)]; + tensor var_646 = matmul(transpose_x = var_646_transpose_x_1, transpose_y = var_646_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_646")]; + tensor new_kv_unnorm_5 = add(x = var_644, y = var_646)[name = tensor("new_kv_unnorm_5")]; + tensor var_648 = const()[name = tensor("op_648"), val = tensor(0x1.4p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_648)[name = tensor("new_scale_5")]; + tensor var_650 = sqrt(x = new_scale_5)[name = tensor("op_650")]; + tensor var_651 = real_div(x = new_kv_unnorm_5, y = var_650)[name = tensor("op_651")]; + tensor var_652_perm_0 = const()[name = tensor("op_652_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_652 = transpose(perm = var_652_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_18, x = var_652)[name = tensor("out_15")]; + tensor var_656 = const()[name = tensor("op_656"), val = tensor([1, 5, 256])]; + tensor out_17 = reshape(shape = var_656, x = out_15)[name = tensor("out_17")]; + tensor var_658 = silu(x = input_97)[name = tensor("op_658")]; + tensor input_99 = mul(x = var_658, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; + tensor var_666_begin_0 = const()[name = tensor("op_666_begin_0"), val = tensor([0, 0, 0])]; + tensor var_666_end_0 = const()[name = tensor("op_666_end_0"), val = tensor([1, 1, 256])]; + tensor var_666_end_mask_0 = const()[name = tensor("op_666_end_mask_0"), val = tensor([true, false, true])]; + tensor var_666 = slice_by_index(begin = var_666_begin_0, end = var_666_end_0, end_mask = var_666_end_mask_0, x = x_15)[name = tensor("op_666")]; + tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 1, 0])]; + tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 16, 256])]; + tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, true])]; + tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = window_25)[name = tensor("op_669")]; + tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; + tensor window_27 = concat(axis = var_27, interleave = window_27_interleave_0, values = (var_669, var_666))[name = tensor("window_27")]; + tensor var_674_begin_0 = const()[name = tensor("op_674_begin_0"), val = tensor([0, 1, 0])]; + tensor var_674_end_0 = const()[name = tensor("op_674_end_0"), val = tensor([1, 2, 256])]; + tensor var_674_end_mask_0 = const()[name = tensor("op_674_end_mask_0"), val = tensor([true, false, true])]; + tensor var_674 = slice_by_index(begin = var_674_begin_0, end = var_674_end_0, end_mask = var_674_end_mask_0, x = x_15)[name = tensor("op_674")]; + tensor var_677_begin_0 = const()[name = tensor("op_677_begin_0"), val = tensor([0, 1, 0])]; + tensor var_677_end_0 = const()[name = tensor("op_677_end_0"), val = tensor([1, 16, 256])]; + tensor var_677_end_mask_0 = const()[name = tensor("op_677_end_mask_0"), val = tensor([true, true, true])]; + tensor var_677 = slice_by_index(begin = var_677_begin_0, end = var_677_end_0, end_mask = var_677_end_mask_0, x = window_27)[name = tensor("op_677")]; + tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; + tensor window_29 = concat(axis = var_27, interleave = window_29_interleave_0, values = (var_677, var_674))[name = tensor("window_29")]; + tensor var_682_begin_0 = const()[name = tensor("op_682_begin_0"), val = tensor([0, 2, 0])]; + tensor var_682_end_0 = const()[name = tensor("op_682_end_0"), val = tensor([1, 3, 256])]; + tensor var_682_end_mask_0 = const()[name = tensor("op_682_end_mask_0"), val = tensor([true, false, true])]; + tensor var_682 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = x_15)[name = tensor("op_682")]; + tensor var_685_begin_0 = const()[name = tensor("op_685_begin_0"), val = tensor([0, 1, 0])]; + tensor var_685_end_0 = const()[name = tensor("op_685_end_0"), val = tensor([1, 16, 256])]; + tensor var_685_end_mask_0 = const()[name = tensor("op_685_end_mask_0"), val = tensor([true, true, true])]; + tensor var_685 = slice_by_index(begin = var_685_begin_0, end = var_685_end_0, end_mask = var_685_end_mask_0, x = window_29)[name = tensor("op_685")]; + tensor window_31_interleave_0 = const()[name = tensor("window_31_interleave_0"), val = tensor(false)]; + tensor window_31 = concat(axis = var_27, interleave = window_31_interleave_0, values = (var_685, var_682))[name = tensor("window_31")]; + tensor var_690_begin_0 = const()[name = tensor("op_690_begin_0"), val = tensor([0, 3, 0])]; + tensor var_690_end_0 = const()[name = tensor("op_690_end_0"), val = tensor([1, 4, 256])]; + tensor var_690_end_mask_0 = const()[name = tensor("op_690_end_mask_0"), val = tensor([true, false, true])]; + tensor var_690 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, x = x_15)[name = tensor("op_690")]; + tensor var_693_begin_0 = const()[name = tensor("op_693_begin_0"), val = tensor([0, 1, 0])]; + tensor var_693_end_0 = const()[name = tensor("op_693_end_0"), val = tensor([1, 16, 256])]; + tensor var_693_end_mask_0 = const()[name = tensor("op_693_end_mask_0"), val = tensor([true, true, true])]; + tensor var_693 = slice_by_index(begin = var_693_begin_0, end = var_693_end_0, end_mask = var_693_end_mask_0, x = window_31)[name = tensor("op_693")]; + tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; + tensor window_33 = concat(axis = var_27, interleave = window_33_interleave_0, values = (var_693, var_690))[name = tensor("window_33")]; + tensor var_698_begin_0 = const()[name = tensor("op_698_begin_0"), val = tensor([0, 4, 0])]; + tensor var_698_end_0 = const()[name = tensor("op_698_end_0"), val = tensor([1, 1, 256])]; + tensor var_698_end_mask_0 = const()[name = tensor("op_698_end_mask_0"), val = tensor([true, true, true])]; + tensor var_698 = slice_by_index(begin = var_698_begin_0, end = var_698_end_0, end_mask = var_698_end_mask_0, x = x_15)[name = tensor("op_698")]; + tensor var_701_begin_0 = const()[name = tensor("op_701_begin_0"), val = tensor([0, 1, 0])]; + tensor var_701_end_0 = const()[name = tensor("op_701_end_0"), val = tensor([1, 16, 256])]; + tensor var_701_end_mask_0 = const()[name = tensor("op_701_end_mask_0"), val = tensor([true, true, true])]; + tensor var_701 = slice_by_index(begin = var_701_begin_0, end = var_701_end_0, end_mask = var_701_end_mask_0, x = window_33)[name = tensor("op_701")]; + tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; + tensor window_35 = concat(axis = var_27, interleave = window_35_interleave_0, values = (var_701, var_698))[name = tensor("window_35")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_24, interleave = input_101_interleave_0, values = (window_27, window_29, window_31, window_33, window_35))[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_726_split_sizes_0 = const()[name = tensor("op_726_split_sizes_0"), val = tensor([256, 256])]; + tensor var_726_axis_0 = const()[name = tensor("op_726_axis_0"), val = tensor(1)]; + tensor var_726_0, tensor var_726_1 = split(axis = var_726_axis_0, split_sizes = var_726_split_sizes_0, x = inputs_23)[name = tensor("op_726")]; + tensor var_728 = sigmoid(x = var_726_1)[name = tensor("op_728")]; + tensor inputs_25 = mul(x = var_726_0, y = var_728)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([5, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([0, -1, 0])]; + tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([5, 16, 256])]; + tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_759 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, x = conv_out_5)[name = tensor("op_759")]; + tensor var_761_perm_0 = const()[name = tensor("op_761_perm_0"), val = tensor([1, 0, 2])]; + tensor var_761 = transpose(perm = var_761_perm_0, x = var_759)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_761)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_784 = const()[name = tensor("op_784"), val = tensor(0x1p-1)]; + tensor var_785 = mul(x = input_119, y = var_784)[name = tensor("op_785")]; + tensor input_121 = add(x = var_785, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_29, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_814 = const()[name = tensor("op_814"), val = tensor(0x1p-1)]; + tensor var_815 = mul(x = input_131, y = var_814)[name = tensor("op_815")]; + tensor input_133 = add(x = var_815, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_29, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_829 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_830 = const()[name = tensor("op_830"), val = tensor([1, 5, 4, 64])]; + tensor var_831 = reshape(shape = var_830, x = var_829)[name = tensor("op_831")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_835 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_836 = const()[name = tensor("op_836"), val = tensor(0x1p-3)]; + tensor var_837 = mul(x = var_835, y = var_836)[name = tensor("op_837")]; + tensor var_838 = const()[name = tensor("op_838"), val = tensor([1, 5, 4, 64])]; + tensor var_839 = reshape(shape = var_838, x = var_837)[name = tensor("op_839")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_843 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_844 = const()[name = tensor("op_844"), val = tensor([1, 5, 4, 64])]; + tensor var_845 = reshape(shape = var_844, x = var_843)[name = tensor("op_845")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_839)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_831)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_855 = const()[name = tensor("op_855"), val = tensor([5, 1])]; + tensor var_856 = reshape(shape = var_855, x = sqrt_s_t_7)[name = tensor("op_856")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_856)[name = tensor("M_7")]; + tensor var_858 = mul(x = qk_7, y = M_7)[name = tensor("op_858")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_845)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_858, y = v_7)[name = tensor("inner_7")]; + tensor var_860_transpose_x_0 = const()[name = tensor("op_860_transpose_x_0"), val = tensor(false)]; + tensor var_860_transpose_y_0 = const()[name = tensor("op_860_transpose_y_0"), val = tensor(false)]; + tensor var_860 = matmul(transpose_x = var_860_transpose_x_0, transpose_y = var_860_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_860")]; + tensor var_861 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_861")]; + tensor var_862 = const()[name = tensor("op_862"), val = tensor([1, 1, 5, 1])]; + tensor var_863 = reshape(shape = var_862, x = var_861)[name = tensor("op_863")]; + tensor cross_7 = mul(x = var_860, y = var_863)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_866 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_866")]; + tensor var_868_transpose_x_1 = const()[name = tensor("op_868_transpose_x_1"), val = tensor(true)]; + tensor var_868_transpose_y_1 = const()[name = tensor("op_868_transpose_y_1"), val = tensor(false)]; + tensor var_868 = matmul(transpose_x = var_868_transpose_x_1, transpose_y = var_868_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_868")]; + tensor new_kv_unnorm_7 = add(x = var_866, y = var_868)[name = tensor("new_kv_unnorm_7")]; + tensor var_870 = const()[name = tensor("op_870"), val = tensor(0x1.4p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_870)[name = tensor("new_scale_7")]; + tensor var_872 = sqrt(x = new_scale_7)[name = tensor("op_872")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_872)[name = tensor("nkv_1")]; + tensor var_874_perm_0 = const()[name = tensor("op_874_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_874 = transpose(perm = var_874_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_18, x = var_874)[name = tensor("out_21")]; + tensor var_878 = const()[name = tensor("op_878"), val = tensor([1, 5, 256])]; + tensor out_23 = reshape(shape = var_878, x = out_21)[name = tensor("out_23")]; + tensor var_880 = silu(x = input_137)[name = tensor("op_880")]; + tensor input_139 = mul(x = var_880, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_37_begin_0 = const()[name = tensor("window_37_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_37_end_0 = const()[name = tensor("window_37_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_37_end_mask_0 = const()[name = tensor("window_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_37_squeeze_mask_0 = const()[name = tensor("window_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_37 = slice_by_index(begin = window_37_begin_0, end = window_37_end_0, end_mask = window_37_end_mask_0, squeeze_mask = window_37_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_37")]; + tensor var_888_begin_0 = const()[name = tensor("op_888_begin_0"), val = tensor([0, 0, 0])]; + tensor var_888_end_0 = const()[name = tensor("op_888_end_0"), val = tensor([1, 1, 256])]; + tensor var_888_end_mask_0 = const()[name = tensor("op_888_end_mask_0"), val = tensor([true, false, true])]; + tensor var_888 = slice_by_index(begin = var_888_begin_0, end = var_888_end_0, end_mask = var_888_end_mask_0, x = x_21)[name = tensor("op_888")]; + tensor var_891_begin_0 = const()[name = tensor("op_891_begin_0"), val = tensor([0, 1, 0])]; + tensor var_891_end_0 = const()[name = tensor("op_891_end_0"), val = tensor([1, 16, 256])]; + tensor var_891_end_mask_0 = const()[name = tensor("op_891_end_mask_0"), val = tensor([true, true, true])]; + tensor var_891 = slice_by_index(begin = var_891_begin_0, end = var_891_end_0, end_mask = var_891_end_mask_0, x = window_37)[name = tensor("op_891")]; + tensor window_39_interleave_0 = const()[name = tensor("window_39_interleave_0"), val = tensor(false)]; + tensor window_39 = concat(axis = var_27, interleave = window_39_interleave_0, values = (var_891, var_888))[name = tensor("window_39")]; + tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 1, 0])]; + tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 2, 256])]; + tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, false, true])]; + tensor var_896 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = x_21)[name = tensor("op_896")]; + tensor var_899_begin_0 = const()[name = tensor("op_899_begin_0"), val = tensor([0, 1, 0])]; + tensor var_899_end_0 = const()[name = tensor("op_899_end_0"), val = tensor([1, 16, 256])]; + tensor var_899_end_mask_0 = const()[name = tensor("op_899_end_mask_0"), val = tensor([true, true, true])]; + tensor var_899 = slice_by_index(begin = var_899_begin_0, end = var_899_end_0, end_mask = var_899_end_mask_0, x = window_39)[name = tensor("op_899")]; + tensor window_41_interleave_0 = const()[name = tensor("window_41_interleave_0"), val = tensor(false)]; + tensor window_41 = concat(axis = var_27, interleave = window_41_interleave_0, values = (var_899, var_896))[name = tensor("window_41")]; + tensor var_904_begin_0 = const()[name = tensor("op_904_begin_0"), val = tensor([0, 2, 0])]; + tensor var_904_end_0 = const()[name = tensor("op_904_end_0"), val = tensor([1, 3, 256])]; + tensor var_904_end_mask_0 = const()[name = tensor("op_904_end_mask_0"), val = tensor([true, false, true])]; + tensor var_904 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = x_21)[name = tensor("op_904")]; + tensor var_907_begin_0 = const()[name = tensor("op_907_begin_0"), val = tensor([0, 1, 0])]; + tensor var_907_end_0 = const()[name = tensor("op_907_end_0"), val = tensor([1, 16, 256])]; + tensor var_907_end_mask_0 = const()[name = tensor("op_907_end_mask_0"), val = tensor([true, true, true])]; + tensor var_907 = slice_by_index(begin = var_907_begin_0, end = var_907_end_0, end_mask = var_907_end_mask_0, x = window_41)[name = tensor("op_907")]; + tensor window_43_interleave_0 = const()[name = tensor("window_43_interleave_0"), val = tensor(false)]; + tensor window_43 = concat(axis = var_27, interleave = window_43_interleave_0, values = (var_907, var_904))[name = tensor("window_43")]; + tensor var_912_begin_0 = const()[name = tensor("op_912_begin_0"), val = tensor([0, 3, 0])]; + tensor var_912_end_0 = const()[name = tensor("op_912_end_0"), val = tensor([1, 4, 256])]; + tensor var_912_end_mask_0 = const()[name = tensor("op_912_end_mask_0"), val = tensor([true, false, true])]; + tensor var_912 = slice_by_index(begin = var_912_begin_0, end = var_912_end_0, end_mask = var_912_end_mask_0, x = x_21)[name = tensor("op_912")]; + tensor var_915_begin_0 = const()[name = tensor("op_915_begin_0"), val = tensor([0, 1, 0])]; + tensor var_915_end_0 = const()[name = tensor("op_915_end_0"), val = tensor([1, 16, 256])]; + tensor var_915_end_mask_0 = const()[name = tensor("op_915_end_mask_0"), val = tensor([true, true, true])]; + tensor var_915 = slice_by_index(begin = var_915_begin_0, end = var_915_end_0, end_mask = var_915_end_mask_0, x = window_43)[name = tensor("op_915")]; + tensor window_45_interleave_0 = const()[name = tensor("window_45_interleave_0"), val = tensor(false)]; + tensor window_45 = concat(axis = var_27, interleave = window_45_interleave_0, values = (var_915, var_912))[name = tensor("window_45")]; + tensor var_920_begin_0 = const()[name = tensor("op_920_begin_0"), val = tensor([0, 4, 0])]; + tensor var_920_end_0 = const()[name = tensor("op_920_end_0"), val = tensor([1, 1, 256])]; + tensor var_920_end_mask_0 = const()[name = tensor("op_920_end_mask_0"), val = tensor([true, true, true])]; + tensor var_920 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = x_21)[name = tensor("op_920")]; + tensor var_923_begin_0 = const()[name = tensor("op_923_begin_0"), val = tensor([0, 1, 0])]; + tensor var_923_end_0 = const()[name = tensor("op_923_end_0"), val = tensor([1, 16, 256])]; + tensor var_923_end_mask_0 = const()[name = tensor("op_923_end_mask_0"), val = tensor([true, true, true])]; + tensor var_923 = slice_by_index(begin = var_923_begin_0, end = var_923_end_0, end_mask = var_923_end_mask_0, x = window_45)[name = tensor("op_923")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_27, interleave = window_interleave_0, values = (var_923, var_920))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_24, interleave = input_141_interleave_0, values = (window_39, window_41, window_43, window_45, window))[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_948_split_sizes_0 = const()[name = tensor("op_948_split_sizes_0"), val = tensor([256, 256])]; + tensor var_948_axis_0 = const()[name = tensor("op_948_axis_0"), val = tensor(1)]; + tensor var_948_0, tensor var_948_1 = split(axis = var_948_axis_0, split_sizes = var_948_split_sizes_0, x = inputs_33)[name = tensor("op_948")]; + tensor var_950 = sigmoid(x = var_948_1)[name = tensor("op_950")]; + tensor inputs_35 = mul(x = var_948_0, y = var_950)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([5, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_981_begin_0 = const()[name = tensor("op_981_begin_0"), val = tensor([0, -1, 0])]; + tensor var_981_end_0 = const()[name = tensor("op_981_end_0"), val = tensor([5, 16, 256])]; + tensor var_981_end_mask_0 = const()[name = tensor("op_981_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_981 = slice_by_index(begin = var_981_begin_0, end = var_981_end_0, end_mask = var_981_end_mask_0, x = conv_out_7)[name = tensor("op_981")]; + tensor var_983_perm_0 = const()[name = tensor("op_983_perm_0"), val = tensor([1, 0, 2])]; + tensor var_983 = transpose(perm = var_983_perm_0, x = var_981)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_983)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_1006 = const()[name = tensor("op_1006"), val = tensor(0x1p-1)]; + tensor var_1007 = mul(x = input_159, y = var_1006)[name = tensor("op_1007")]; + tensor input_161 = add(x = var_1007, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_29, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_21, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1025_begin_0 = const()[name = tensor("op_1025_begin_0"), val = tensor([0, 0, 5])]; + tensor var_1025_end_0 = const()[name = tensor("op_1025_end_0"), val = tensor([1, 256, 23])]; + tensor var_1025_end_mask_0 = const()[name = tensor("op_1025_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1025_begin_0, end = var_1025_end_0, end_mask = var_1025_end_mask_0, x = cat)[name = tensor("op_1025")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1028 = reduce_l2_norm(axes = var_1027, keep_dims = var_30, x = input_163)[name = tensor("op_1028")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_1028)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_1032_axis_0 = const()[name = tensor("op_1032_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1032_axis_0, values = (var_207, var_429, var_651, nkv_1))[name = tensor("op_1032")]; + tensor var_1034_axis_0 = const()[name = tensor("op_1034_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1034_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1034")]; + tensor var_1036_axis_0 = const()[name = tensor("op_1036_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1036_axis_0, values = (window_11, window_23, window_35, window))[name = tensor("op_1036")]; + tensor var_1045 = const()[name = tensor("op_1045"), val = tensor(0x1.5798eep-27)]; + tensor var_1050 = const()[name = tensor("op_1050"), val = tensor(0x1.4f8b58p-17)]; + tensor var_1052 = const()[name = tensor("op_1052"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_1053 = const()[name = tensor("op_1053"), val = tensor(true)]; + tensor var_1055 = const()[name = tensor("op_1055"), val = tensor(0x1p+0)]; + tensor var_1059 = const()[name = tensor("op_1059"), val = tensor(-1)]; + tensor var_1065 = const()[name = tensor("op_1065"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395712)))]; + tensor var_1127_axes_0 = const()[name = tensor("op_1127_axes_0"), val = tensor([2])]; + tensor var_1127 = expand_dims(axes = var_1127_axes_0, x = emb)[name = tensor("op_1127")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 6, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1127)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_1059, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1135_perm_0 = const()[name = tensor("op_1135_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1139 = const()[name = tensor("op_1139"), val = tensor([6, 5, 256])]; + tensor var_1135 = transpose(perm = var_1135_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1139, x = var_1135)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 6, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1147 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([6, 5, 4, 64])]; + tensor var_1149 = reshape(shape = var_1148, x = var_1147)[name = tensor("op_1149")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1153 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor(0x1p-3)]; + tensor var_1155 = mul(x = var_1153, y = var_1154)[name = tensor("op_1155")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([6, 5, 4, 64])]; + tensor var_1157 = reshape(shape = var_1156, x = var_1155)[name = tensor("op_1157")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1161 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor([6, 5, 4, 64])]; + tensor var_1163 = reshape(shape = var_1162, x = var_1161)[name = tensor("op_1163")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_1065, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_1055, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1157)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1149)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 5])]; + tensor var_1176 = reshape(shape = var_1175, x = valid_mask)[name = tensor("op_1176")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1176)[name = tensor("causal_with_valid_1")]; + tensor var_1178 = const()[name = tensor("op_1178"), val = tensor([5, 1])]; + tensor var_1179 = reshape(shape = var_1178, x = sqrt_s_t_9)[name = tensor("op_1179")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1179)[name = tensor("M_9")]; + tensor var_1181 = mul(x = qk_9, y = M_9)[name = tensor("op_1181")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1163)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1181, y = v_9)[name = tensor("inner_9")]; + tensor var_1183_transpose_x_0 = const()[name = tensor("op_1183_transpose_x_0"), val = tensor(false)]; + tensor var_1183_transpose_y_0 = const()[name = tensor("op_1183_transpose_y_0"), val = tensor(false)]; + tensor var_1183 = matmul(transpose_x = var_1183_transpose_x_0, transpose_y = var_1183_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1183")]; + tensor var_1184 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1184")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 5, 1])]; + tensor var_1186 = reshape(shape = var_1185, x = var_1184)[name = tensor("op_1186")]; + tensor cross_9 = mul(x = var_1183, y = var_1186)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1, 1, 5, 1])]; + tensor var_1190 = reshape(shape = var_1189, x = valid_mask)[name = tensor("op_1190")]; + tensor v_masked_1 = mul(x = v_9, y = var_1190)[name = tensor("v_masked_1")]; + tensor var_1192 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1192")]; + tensor var_1194_transpose_x_1 = const()[name = tensor("op_1194_transpose_x_1"), val = tensor(true)]; + tensor var_1194_transpose_y_1 = const()[name = tensor("op_1194_transpose_y_1"), val = tensor(false)]; + tensor var_1194 = matmul(transpose_x = var_1194_transpose_x_1, transpose_y = var_1194_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1194")]; + tensor new_kv_unnorm_9 = add(x = var_1192, y = var_1194)[name = tensor("new_kv_unnorm_9")]; + tensor var_1196_keep_dims_0 = const()[name = tensor("op_1196_keep_dims_0"), val = tensor(false)]; + tensor var_1196 = reduce_sum(keep_dims = var_1196_keep_dims_0, x = valid_mask)[name = tensor("op_1196")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1])]; + tensor var_1198 = reshape(shape = var_1197, x = var_1196)[name = tensor("op_1198")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1198)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_1055, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1202 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1202")]; + tensor var_1203_perm_0 = const()[name = tensor("op_1203_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1203 = transpose(perm = var_1203_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1052, x = var_1203)[name = tensor("out_27")]; + tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([6, 5, 256])]; + tensor out_29 = reshape(shape = var_1207, x = out_27)[name = tensor("out_29")]; + tensor var_1209 = silu(x = input_169)[name = tensor("op_1209")]; + tensor input_171 = mul(x = var_1209, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1050, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1219 = const()[name = tensor("op_1219"), val = tensor([1, 6, 5, 256])]; + tensor var_1220 = reshape(shape = var_1219, x = xt_1)[name = tensor("op_1220")]; + tensor var_1221_perm_0 = const()[name = tensor("op_1221_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([5, 6, 256])]; + tensor var_1221 = transpose(perm = var_1221_perm_0, x = var_1220)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1224, x = var_1221)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1247 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([6, 5, 3, 256])]; + tensor var_1249 = reshape(shape = concat_1, x = var_1247)[name = tensor("op_1249")]; + tensor var_1250_axes_0 = const()[name = tensor("op_1250_axes_0"), val = tensor([0])]; + tensor var_1250 = expand_dims(axes = var_1250_axes_0, x = var_1249)[name = tensor("op_1250")]; + tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1252_axes_0 = const()[name = tensor("op_1252_axes_0"), val = tensor([-2])]; + tensor var_1251 = transpose(perm = var_1251_perm_0, x = var_1250)[name = tensor("transpose_21")]; + tensor var_1252 = squeeze(axes = var_1252_axes_0, x = var_1251)[name = tensor("op_1252")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 6, 5, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1252)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 6, 5, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1252)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 6, 5, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1252)[name = tensor("v_11")]; + tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([6, 20, 64])]; + tensor var_1261 = reshape(shape = var_1260, x = q_11)[name = tensor("op_1261")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1267 = const()[name = tensor("op_1267"), val = tensor([6, 20, 64])]; + tensor var_1268 = reshape(shape = var_1267, x = k_11)[name = tensor("op_1268")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([6, 20, 64])]; + tensor var_1275 = reshape(shape = var_1274, x = v_11)[name = tensor("op_1275")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([5, 4, 6, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1261)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1278, x = q_13)[name = tensor("q_15")]; + tensor var_1280 = const()[name = tensor("op_1280"), val = tensor([5, 4, 6, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1268)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1280, x = k_13)[name = tensor("k_15")]; + tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([5, 4, 6, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1275)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1282, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1285 = const()[name = tensor("op_1285"), val = tensor([2, 0, 1, 3])]; + tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([30, 256])]; + tensor var_1286 = transpose(perm = var_1285, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1290, x = var_1286)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([6, 5, 256])]; + tensor attn_output_7 = reshape(shape = var_1294, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1050, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1050, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 5, 6, 256])]; + tensor x_31 = reshape(shape = var_1314, x = xt_3)[name = tensor("x_31")]; + tensor var_1316_perm_0 = const()[name = tensor("op_1316_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1320 = const()[name = tensor("op_1320"), val = tensor([6, 5, 256])]; + tensor var_1316 = transpose(perm = var_1316_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1320, x = var_1316)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 6, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1328 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([6, 5, 4, 64])]; + tensor var_1330 = reshape(shape = var_1329, x = var_1328)[name = tensor("op_1330")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1334 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor(0x1p-3)]; + tensor var_1336 = mul(x = var_1334, y = var_1335)[name = tensor("op_1336")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([6, 5, 4, 64])]; + tensor var_1338 = reshape(shape = var_1337, x = var_1336)[name = tensor("op_1338")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1342 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([6, 5, 4, 64])]; + tensor var_1344 = reshape(shape = var_1343, x = var_1342)[name = tensor("op_1344")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_1055, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1338)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1330)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1359 = const()[name = tensor("op_1359"), val = tensor([5, 1])]; + tensor var_1360 = reshape(shape = var_1359, x = sqrt_s_t)[name = tensor("op_1360")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1360)[name = tensor("M")]; + tensor var_1362 = mul(x = qk, y = M)[name = tensor("op_1362")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1344)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1362, y = v_17)[name = tensor("inner")]; + tensor var_1364_transpose_x_0 = const()[name = tensor("op_1364_transpose_x_0"), val = tensor(false)]; + tensor var_1364_transpose_y_0 = const()[name = tensor("op_1364_transpose_y_0"), val = tensor(false)]; + tensor var_1364 = matmul(transpose_x = var_1364_transpose_x_0, transpose_y = var_1364_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1364")]; + tensor var_1365 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1365")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 5, 1])]; + tensor var_1367 = reshape(shape = var_1366, x = var_1365)[name = tensor("op_1367")]; + tensor cross = mul(x = var_1364, y = var_1367)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1190)[name = tensor("v_masked")]; + tensor var_1373 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1373")]; + tensor var_1375_transpose_x_1 = const()[name = tensor("op_1375_transpose_x_1"), val = tensor(true)]; + tensor var_1375_transpose_y_1 = const()[name = tensor("op_1375_transpose_y_1"), val = tensor(false)]; + tensor var_1375 = matmul(transpose_x = var_1375_transpose_x_1, transpose_y = var_1375_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1375")]; + tensor new_kv_unnorm = add(x = var_1373, y = var_1375)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1198)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_1055, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1384_perm_0 = const()[name = tensor("op_1384_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1384 = transpose(perm = var_1384_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1052, x = var_1384)[name = tensor("out_33")]; + tensor var_1388 = const()[name = tensor("op_1388"), val = tensor([6, 5, 256])]; + tensor out = reshape(shape = var_1388, x = out_33)[name = tensor("out")]; + tensor var_1390 = silu(x = input_187)[name = tensor("op_1390")]; + tensor input_189 = mul(x = var_1390, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1050, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 6, 5, 256])]; + tensor var_1401 = reshape(shape = var_1400, x = xt_5)[name = tensor("op_1401")]; + tensor var_1402_perm_0 = const()[name = tensor("op_1402_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1405 = const()[name = tensor("op_1405"), val = tensor([5, 6, 256])]; + tensor var_1402 = transpose(perm = var_1402_perm_0, x = var_1401)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1405, x = var_1402)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1428 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([6, 5, 3, 256])]; + tensor var_1430 = reshape(shape = concat_2, x = var_1428)[name = tensor("op_1430")]; + tensor var_1431_axes_0 = const()[name = tensor("op_1431_axes_0"), val = tensor([0])]; + tensor var_1431 = expand_dims(axes = var_1431_axes_0, x = var_1430)[name = tensor("op_1431")]; + tensor var_1432_perm_0 = const()[name = tensor("op_1432_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1433_axes_0 = const()[name = tensor("op_1433_axes_0"), val = tensor([-2])]; + tensor var_1432 = transpose(perm = var_1432_perm_0, x = var_1431)[name = tensor("transpose_8")]; + tensor var_1433 = squeeze(axes = var_1433_axes_0, x = var_1432)[name = tensor("op_1433")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 6, 5, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1433)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 6, 5, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1433)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 6, 5, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1433)[name = tensor("v_19")]; + tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([6, 20, 64])]; + tensor var_1442 = reshape(shape = var_1441, x = q_19)[name = tensor("op_1442")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([6, 20, 64])]; + tensor var_1449 = reshape(shape = var_1448, x = k_19)[name = tensor("op_1449")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([6, 20, 64])]; + tensor var_1456 = reshape(shape = var_1455, x = v_19)[name = tensor("op_1456")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([5, 4, 6, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1442)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1459, x = q_21)[name = tensor("q")]; + tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([5, 4, 6, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1449)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1461, x = k_21)[name = tensor("k")]; + tensor var_1463 = const()[name = tensor("op_1463"), val = tensor([5, 4, 6, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1456)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1463, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1466 = const()[name = tensor("op_1466"), val = tensor([2, 0, 1, 3])]; + tensor var_1471 = const()[name = tensor("op_1471"), val = tensor([30, 256])]; + tensor var_1467 = transpose(perm = var_1466, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1471, x = var_1467)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1475 = const()[name = tensor("op_1475"), val = tensor([6, 5, 256])]; + tensor attn_output = reshape(shape = var_1475, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1050, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1050, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1495 = const()[name = tensor("op_1495"), val = tensor([1, 5, 6, 256])]; + tensor input = reshape(shape = var_1495, x = xt)[name = tensor("input")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([-1])]; + tensor var_1498 = reduce_l2_norm(axes = var_1497, keep_dims = var_1053, x = input)[name = tensor("op_1498")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_1045, beta = const_42, x = var_1498)[name = tensor("clip_5")]; + tensor var_1500 = real_div(x = input, y = clip_5)[name = tensor("op_1500")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([5, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([5, 256, 6])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1500)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 5, 6])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 5, 5])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1504")]; + tensor var_1506_axis_0 = const()[name = tensor("op_1506_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1506_axis_0, values = (var_1202, nkv))[name = tensor("op_1506")]; + tensor var_1508_axis_0 = const()[name = tensor("op_1508_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1508_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1508")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/weights/weight.bin b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..69ad971869367fee2a996c16cc8a622087e55bbb --- /dev/null +++ b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97952c32e939e275869cf202b0079b5563cf071ca90b2df10a123d1f56e702c5 +size 44426496 diff --git a/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..c2635f870434e5d00bdf7febc221c1084a938f1b --- /dev/null +++ b/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb1fc846d5fe66b99cc3d94f2f27dc499fd4bcd04080e23e0297dd02f0f61f7b +size 196611 diff --git a/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..69ad971869367fee2a996c16cc8a622087e55bbb --- /dev/null +++ b/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97952c32e939e275869cf202b0079b5563cf071ca90b2df10a123d1f56e702c5 +size 44426496 diff --git a/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Manifest.json b/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..719daafccc33381f00ca4380bee97b50ce8c7019 --- /dev/null +++ b/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "BFFA710C-2BC9-41A2-80C8-E692BA4158D2": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "F337566C-1A74-4DED-8712-B263AAD1B3CF": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "F337566C-1A74-4DED-8712-B263AAD1B3CF" +} diff --git a/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/analytics/coremldata.bin b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d14d2473e765f039603cf751a34e8895b6d8f152 --- /dev/null +++ b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:162c9b06e4452aa666026e5cf6d51ebd396faac77145f39864e8e1fee2ec569e +size 243 diff --git a/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/coremldata.bin b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..4f3888dde2e2edab7506a40b902b161ff6a88031 --- /dev/null +++ b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e2c9fc79e93736f5e7e4017f5282df7b689fac77791ab360645dd0732ca2af9 +size 1301 diff --git a/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/metadata.json b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..c0c60029c9245e143c373a94558005a27729a798 --- /dev/null +++ b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND CALLHOME streaming diarizer (pipeline, T=1, max_speakers=7)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 7)", + "shortDescription" : "", + "shape" : "[1, 1, 7]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 9 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 9, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 66, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 36, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 10, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 46 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 345)", + "shortDescription" : "", + "shape" : "[1, 1, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 9 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 9, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"ch\", \"model_label\": \"CALLHOME\", \"variant\": \"pipeline\", \"chunk_size\": 1, \"step_duration_ms\": 100, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 7, \"max_nspks\": 9, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_ch_100ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/model.mil b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..c6564bd8478dcda29a396e2ab8dcb7eda1e4a19b --- /dev/null +++ b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/model.mil @@ -0,0 +1,1184 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0]])]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 1, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 1, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 1, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 1, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+0)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 1, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, x_3))[name = tensor("window_3")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = window_3)[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_249_split_sizes_0 = const()[name = tensor("op_249_split_sizes_0"), val = tensor([256, 256])]; + tensor var_249_axis_0 = const()[name = tensor("op_249_axis_0"), val = tensor(1)]; + tensor var_249_0, tensor var_249_1 = split(axis = var_249_axis_0, split_sizes = var_249_split_sizes_0, x = inputs_3)[name = tensor("op_249")]; + tensor var_251 = sigmoid(x = var_249_1)[name = tensor("op_251")]; + tensor inputs_5 = mul(x = var_249_0, y = var_251)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([1, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_282_begin_0 = const()[name = tensor("op_282_begin_0"), val = tensor([0, -1, 0])]; + tensor var_282_end_0 = const()[name = tensor("op_282_end_0"), val = tensor([1, 16, 256])]; + tensor var_282_end_mask_0 = const()[name = tensor("op_282_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_282 = slice_by_index(begin = var_282_begin_0, end = var_282_end_0, end_mask = var_282_end_mask_0, x = conv_out_1)[name = tensor("op_282")]; + tensor var_284_perm_0 = const()[name = tensor("op_284_perm_0"), val = tensor([1, 0, 2])]; + tensor var_284 = transpose(perm = var_284_perm_0, x = var_282)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_284)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_307 = const()[name = tensor("op_307"), val = tensor(0x1p-1)]; + tensor var_308 = mul(x = input_39, y = var_307)[name = tensor("op_308")]; + tensor input_41 = add(x = var_308, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_337 = const()[name = tensor("op_337"), val = tensor(0x1p-1)]; + tensor var_338 = mul(x = input_51, y = var_337)[name = tensor("op_338")]; + tensor input_53 = add(x = var_338, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_352 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_353 = const()[name = tensor("op_353"), val = tensor([1, 1, 4, 64])]; + tensor var_354 = reshape(shape = var_353, x = var_352)[name = tensor("op_354")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_358 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_359 = const()[name = tensor("op_359"), val = tensor(0x1p-3)]; + tensor var_360 = mul(x = var_358, y = var_359)[name = tensor("op_360")]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 1, 4, 64])]; + tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_366 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_367 = const()[name = tensor("op_367"), val = tensor([1, 1, 4, 64])]; + tensor var_368 = reshape(shape = var_367, x = var_366)[name = tensor("op_368")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_362)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_354)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 1])]; + tensor var_379 = reshape(shape = var_378, x = sqrt_s_t_3)[name = tensor("op_379")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_379)[name = tensor("M_3")]; + tensor var_381 = mul(x = qk_3, y = M_3)[name = tensor("op_381")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_368)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_381, y = v_3)[name = tensor("inner_3")]; + tensor var_383_transpose_x_0 = const()[name = tensor("op_383_transpose_x_0"), val = tensor(false)]; + tensor var_383_transpose_y_0 = const()[name = tensor("op_383_transpose_y_0"), val = tensor(false)]; + tensor var_383 = matmul(transpose_x = var_383_transpose_x_0, transpose_y = var_383_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_383")]; + tensor var_384 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_384")]; + tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 1, 1, 1])]; + tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; + tensor cross_3 = mul(x = var_383, y = var_386)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_389 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_389")]; + tensor var_391_transpose_x_1 = const()[name = tensor("op_391_transpose_x_1"), val = tensor(true)]; + tensor var_391_transpose_y_1 = const()[name = tensor("op_391_transpose_y_1"), val = tensor(false)]; + tensor var_391 = matmul(transpose_x = var_391_transpose_x_1, transpose_y = var_391_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_391")]; + tensor new_kv_unnorm_3 = add(x = var_389, y = var_391)[name = tensor("new_kv_unnorm_3")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor(0x1p+0)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_393)[name = tensor("new_scale_3")]; + tensor var_395 = sqrt(x = new_scale_3)[name = tensor("op_395")]; + tensor var_396 = real_div(x = new_kv_unnorm_3, y = var_395)[name = tensor("op_396")]; + tensor var_397_perm_0 = const()[name = tensor("op_397_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_397 = transpose(perm = var_397_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_397)[name = tensor("out_9")]; + tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 256])]; + tensor out_11 = reshape(shape = var_401, x = out_9)[name = tensor("out_11")]; + tensor var_403 = silu(x = input_57)[name = tensor("op_403")]; + tensor input_59 = mul(x = var_403, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_5_begin_0 = const()[name = tensor("window_5_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_5_end_0 = const()[name = tensor("window_5_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_5_end_mask_0 = const()[name = tensor("window_5_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_5_squeeze_mask_0 = const()[name = tensor("window_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_5 = slice_by_index(begin = window_5_begin_0, end = window_5_end_0, end_mask = window_5_end_mask_0, squeeze_mask = window_5_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_5")]; + tensor var_414_begin_0 = const()[name = tensor("op_414_begin_0"), val = tensor([0, 1, 0])]; + tensor var_414_end_0 = const()[name = tensor("op_414_end_0"), val = tensor([1, 16, 256])]; + tensor var_414_end_mask_0 = const()[name = tensor("op_414_end_mask_0"), val = tensor([true, true, true])]; + tensor var_414 = slice_by_index(begin = var_414_begin_0, end = var_414_end_0, end_mask = var_414_end_mask_0, x = window_5)[name = tensor("op_414")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_414, x_9))[name = tensor("window_7")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = window_7)[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_439_split_sizes_0 = const()[name = tensor("op_439_split_sizes_0"), val = tensor([256, 256])]; + tensor var_439_axis_0 = const()[name = tensor("op_439_axis_0"), val = tensor(1)]; + tensor var_439_0, tensor var_439_1 = split(axis = var_439_axis_0, split_sizes = var_439_split_sizes_0, x = inputs_13)[name = tensor("op_439")]; + tensor var_441 = sigmoid(x = var_439_1)[name = tensor("op_441")]; + tensor inputs_15 = mul(x = var_439_0, y = var_441)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([1, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_472_begin_0 = const()[name = tensor("op_472_begin_0"), val = tensor([0, -1, 0])]; + tensor var_472_end_0 = const()[name = tensor("op_472_end_0"), val = tensor([1, 16, 256])]; + tensor var_472_end_mask_0 = const()[name = tensor("op_472_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_472 = slice_by_index(begin = var_472_begin_0, end = var_472_end_0, end_mask = var_472_end_mask_0, x = conv_out_3)[name = tensor("op_472")]; + tensor var_474_perm_0 = const()[name = tensor("op_474_perm_0"), val = tensor([1, 0, 2])]; + tensor var_474 = transpose(perm = var_474_perm_0, x = var_472)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_474)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_497 = const()[name = tensor("op_497"), val = tensor(0x1p-1)]; + tensor var_498 = mul(x = input_79, y = var_497)[name = tensor("op_498")]; + tensor input_81 = add(x = var_498, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_527 = const()[name = tensor("op_527"), val = tensor(0x1p-1)]; + tensor var_528 = mul(x = input_91, y = var_527)[name = tensor("op_528")]; + tensor input_93 = add(x = var_528, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_542 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 1, 4, 64])]; + tensor var_544 = reshape(shape = var_543, x = var_542)[name = tensor("op_544")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_548 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_549 = const()[name = tensor("op_549"), val = tensor(0x1p-3)]; + tensor var_550 = mul(x = var_548, y = var_549)[name = tensor("op_550")]; + tensor var_551 = const()[name = tensor("op_551"), val = tensor([1, 1, 4, 64])]; + tensor var_552 = reshape(shape = var_551, x = var_550)[name = tensor("op_552")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_556 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_557 = const()[name = tensor("op_557"), val = tensor([1, 1, 4, 64])]; + tensor var_558 = reshape(shape = var_557, x = var_556)[name = tensor("op_558")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_552)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_544)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_568 = const()[name = tensor("op_568"), val = tensor([1, 1])]; + tensor var_569 = reshape(shape = var_568, x = sqrt_s_t_5)[name = tensor("op_569")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_569)[name = tensor("M_5")]; + tensor var_571 = mul(x = qk_5, y = M_5)[name = tensor("op_571")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_558)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_571, y = v_5)[name = tensor("inner_5")]; + tensor var_573_transpose_x_0 = const()[name = tensor("op_573_transpose_x_0"), val = tensor(false)]; + tensor var_573_transpose_y_0 = const()[name = tensor("op_573_transpose_y_0"), val = tensor(false)]; + tensor var_573 = matmul(transpose_x = var_573_transpose_x_0, transpose_y = var_573_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_573")]; + tensor var_574 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_574")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 1, 1, 1])]; + tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; + tensor cross_5 = mul(x = var_573, y = var_576)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_579 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_579")]; + tensor var_581_transpose_x_1 = const()[name = tensor("op_581_transpose_x_1"), val = tensor(true)]; + tensor var_581_transpose_y_1 = const()[name = tensor("op_581_transpose_y_1"), val = tensor(false)]; + tensor var_581 = matmul(transpose_x = var_581_transpose_x_1, transpose_y = var_581_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_581")]; + tensor new_kv_unnorm_5 = add(x = var_579, y = var_581)[name = tensor("new_kv_unnorm_5")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor(0x1p+0)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_583)[name = tensor("new_scale_5")]; + tensor var_585 = sqrt(x = new_scale_5)[name = tensor("op_585")]; + tensor var_586 = real_div(x = new_kv_unnorm_5, y = var_585)[name = tensor("op_586")]; + tensor var_587_perm_0 = const()[name = tensor("op_587_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_587 = transpose(perm = var_587_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_587)[name = tensor("out_15")]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 256])]; + tensor out_17 = reshape(shape = var_591, x = out_15)[name = tensor("out_17")]; + tensor var_593 = silu(x = input_97)[name = tensor("op_593")]; + tensor input_99 = mul(x = var_593, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; + tensor var_604_begin_0 = const()[name = tensor("op_604_begin_0"), val = tensor([0, 1, 0])]; + tensor var_604_end_0 = const()[name = tensor("op_604_end_0"), val = tensor([1, 16, 256])]; + tensor var_604_end_mask_0 = const()[name = tensor("op_604_end_mask_0"), val = tensor([true, true, true])]; + tensor var_604 = slice_by_index(begin = var_604_begin_0, end = var_604_end_0, end_mask = var_604_end_mask_0, x = window_9)[name = tensor("op_604")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_604, x_15))[name = tensor("window_11")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = window_11)[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_629_split_sizes_0 = const()[name = tensor("op_629_split_sizes_0"), val = tensor([256, 256])]; + tensor var_629_axis_0 = const()[name = tensor("op_629_axis_0"), val = tensor(1)]; + tensor var_629_0, tensor var_629_1 = split(axis = var_629_axis_0, split_sizes = var_629_split_sizes_0, x = inputs_23)[name = tensor("op_629")]; + tensor var_631 = sigmoid(x = var_629_1)[name = tensor("op_631")]; + tensor inputs_25 = mul(x = var_629_0, y = var_631)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([1, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_662_begin_0 = const()[name = tensor("op_662_begin_0"), val = tensor([0, -1, 0])]; + tensor var_662_end_0 = const()[name = tensor("op_662_end_0"), val = tensor([1, 16, 256])]; + tensor var_662_end_mask_0 = const()[name = tensor("op_662_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_662 = slice_by_index(begin = var_662_begin_0, end = var_662_end_0, end_mask = var_662_end_mask_0, x = conv_out_5)[name = tensor("op_662")]; + tensor var_664_perm_0 = const()[name = tensor("op_664_perm_0"), val = tensor([1, 0, 2])]; + tensor var_664 = transpose(perm = var_664_perm_0, x = var_662)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_664)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_687 = const()[name = tensor("op_687"), val = tensor(0x1p-1)]; + tensor var_688 = mul(x = input_119, y = var_687)[name = tensor("op_688")]; + tensor input_121 = add(x = var_688, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_717 = const()[name = tensor("op_717"), val = tensor(0x1p-1)]; + tensor var_718 = mul(x = input_131, y = var_717)[name = tensor("op_718")]; + tensor input_133 = add(x = var_718, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_732 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_733 = const()[name = tensor("op_733"), val = tensor([1, 1, 4, 64])]; + tensor var_734 = reshape(shape = var_733, x = var_732)[name = tensor("op_734")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_738 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_739 = const()[name = tensor("op_739"), val = tensor(0x1p-3)]; + tensor var_740 = mul(x = var_738, y = var_739)[name = tensor("op_740")]; + tensor var_741 = const()[name = tensor("op_741"), val = tensor([1, 1, 4, 64])]; + tensor var_742 = reshape(shape = var_741, x = var_740)[name = tensor("op_742")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_746 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 1, 4, 64])]; + tensor var_748 = reshape(shape = var_747, x = var_746)[name = tensor("op_748")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_742)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_734)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_758 = const()[name = tensor("op_758"), val = tensor([1, 1])]; + tensor var_759 = reshape(shape = var_758, x = sqrt_s_t_7)[name = tensor("op_759")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_759)[name = tensor("M_7")]; + tensor var_761 = mul(x = qk_7, y = M_7)[name = tensor("op_761")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_748)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_761, y = v_7)[name = tensor("inner_7")]; + tensor var_763_transpose_x_0 = const()[name = tensor("op_763_transpose_x_0"), val = tensor(false)]; + tensor var_763_transpose_y_0 = const()[name = tensor("op_763_transpose_y_0"), val = tensor(false)]; + tensor var_763 = matmul(transpose_x = var_763_transpose_x_0, transpose_y = var_763_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_763")]; + tensor var_764 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_764")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 1, 1, 1])]; + tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; + tensor cross_7 = mul(x = var_763, y = var_766)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_769 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_769")]; + tensor var_771_transpose_x_1 = const()[name = tensor("op_771_transpose_x_1"), val = tensor(true)]; + tensor var_771_transpose_y_1 = const()[name = tensor("op_771_transpose_y_1"), val = tensor(false)]; + tensor var_771 = matmul(transpose_x = var_771_transpose_x_1, transpose_y = var_771_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_771")]; + tensor new_kv_unnorm_7 = add(x = var_769, y = var_771)[name = tensor("new_kv_unnorm_7")]; + tensor var_773 = const()[name = tensor("op_773"), val = tensor(0x1p+0)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_773)[name = tensor("new_scale_7")]; + tensor var_775 = sqrt(x = new_scale_7)[name = tensor("op_775")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_775)[name = tensor("nkv_1")]; + tensor var_777_perm_0 = const()[name = tensor("op_777_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_777 = transpose(perm = var_777_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_777)[name = tensor("out_21")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 1, 256])]; + tensor out_23 = reshape(shape = var_781, x = out_21)[name = tensor("out_23")]; + tensor var_783 = silu(x = input_137)[name = tensor("op_783")]; + tensor input_139 = mul(x = var_783, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; + tensor var_794_begin_0 = const()[name = tensor("op_794_begin_0"), val = tensor([0, 1, 0])]; + tensor var_794_end_0 = const()[name = tensor("op_794_end_0"), val = tensor([1, 16, 256])]; + tensor var_794_end_mask_0 = const()[name = tensor("op_794_end_mask_0"), val = tensor([true, true, true])]; + tensor var_794 = slice_by_index(begin = var_794_begin_0, end = var_794_end_0, end_mask = var_794_end_mask_0, x = window_13)[name = tensor("op_794")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_794, x_21))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = window)[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_819_split_sizes_0 = const()[name = tensor("op_819_split_sizes_0"), val = tensor([256, 256])]; + tensor var_819_axis_0 = const()[name = tensor("op_819_axis_0"), val = tensor(1)]; + tensor var_819_0, tensor var_819_1 = split(axis = var_819_axis_0, split_sizes = var_819_split_sizes_0, x = inputs_33)[name = tensor("op_819")]; + tensor var_821 = sigmoid(x = var_819_1)[name = tensor("op_821")]; + tensor inputs_35 = mul(x = var_819_0, y = var_821)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([1, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_852_begin_0 = const()[name = tensor("op_852_begin_0"), val = tensor([0, -1, 0])]; + tensor var_852_end_0 = const()[name = tensor("op_852_end_0"), val = tensor([1, 16, 256])]; + tensor var_852_end_mask_0 = const()[name = tensor("op_852_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_852 = slice_by_index(begin = var_852_begin_0, end = var_852_end_0, end_mask = var_852_end_mask_0, x = conv_out_7)[name = tensor("op_852")]; + tensor var_854_perm_0 = const()[name = tensor("op_854_perm_0"), val = tensor([1, 0, 2])]; + tensor var_854 = transpose(perm = var_854_perm_0, x = var_852)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_854)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_877 = const()[name = tensor("op_877"), val = tensor(0x1p-1)]; + tensor var_878 = mul(x = input_159, y = var_877)[name = tensor("op_878")]; + tensor input_161 = add(x = var_878, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 0, 1])]; + tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 256, 19])]; + tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = cat)[name = tensor("op_896")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_898 = const()[name = tensor("op_898"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_899 = reduce_l2_norm(axes = var_898, keep_dims = var_29, x = input_163)[name = tensor("op_899")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_899)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_903_axis_0 = const()[name = tensor("op_903_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_903_axis_0, values = (var_206, var_396, var_586, nkv_1))[name = tensor("op_903")]; + tensor var_905_axis_0 = const()[name = tensor("op_905_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_905_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_905")]; + tensor var_907_axis_0 = const()[name = tensor("op_907_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_907_axis_0, values = (window_3, window_7, window_11, window))[name = tensor("op_907")]; + tensor var_916 = const()[name = tensor("op_916"), val = tensor(0x1.5798eep-27)]; + tensor var_921 = const()[name = tensor("op_921"), val = tensor(0x1.4f8b58p-17)]; + tensor var_923 = const()[name = tensor("op_923"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_924 = const()[name = tensor("op_924"), val = tensor(true)]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor(0x1p+0)]; + tensor var_930 = const()[name = tensor("op_930"), val = tensor(-1)]; + tensor var_936 = const()[name = tensor("op_936"), val = tensor(0)]; + tensor var_993 = const()[name = tensor("op_993"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_998_axes_0 = const()[name = tensor("op_998_axes_0"), val = tensor([2])]; + tensor var_998 = expand_dims(axes = var_998_axes_0, x = emb)[name = tensor("op_998")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 9, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_998)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_930, interleave = input_165_interleave_0, values = (emb_exp, var_993))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1006_perm_0 = const()[name = tensor("op_1006_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([9, 1, 256])]; + tensor var_1006 = transpose(perm = var_1006_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1010, x = var_1006)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 9, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1018 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1019 = const()[name = tensor("op_1019"), val = tensor([9, 1, 4, 64])]; + tensor var_1020 = reshape(shape = var_1019, x = var_1018)[name = tensor("op_1020")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1024 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1025 = const()[name = tensor("op_1025"), val = tensor(0x1p-3)]; + tensor var_1026 = mul(x = var_1024, y = var_1025)[name = tensor("op_1026")]; + tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([9, 1, 4, 64])]; + tensor var_1028 = reshape(shape = var_1027, x = var_1026)[name = tensor("op_1028")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1032 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1033 = const()[name = tensor("op_1033"), val = tensor([9, 1, 4, 64])]; + tensor var_1034 = reshape(shape = var_1033, x = var_1032)[name = tensor("op_1034")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_936, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_926, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1028)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1020)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 1])]; + tensor var_1047 = reshape(shape = var_1046, x = valid_mask)[name = tensor("op_1047")]; + tensor var_1049 = const()[name = tensor("op_1049"), val = tensor([1, 1])]; + tensor var_1050 = reshape(shape = var_1049, x = sqrt_s_t_9)[name = tensor("op_1050")]; + tensor M_9 = real_div(x = var_1047, y = var_1050)[name = tensor("M_9")]; + tensor var_1052 = mul(x = qk_9, y = M_9)[name = tensor("op_1052")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1034)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1052, y = v_9)[name = tensor("inner_9")]; + tensor var_1054_transpose_x_0 = const()[name = tensor("op_1054_transpose_x_0"), val = tensor(false)]; + tensor var_1054_transpose_y_0 = const()[name = tensor("op_1054_transpose_y_0"), val = tensor(false)]; + tensor var_1054 = matmul(transpose_x = var_1054_transpose_x_0, transpose_y = var_1054_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1054")]; + tensor var_1055 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1055")]; + tensor var_1056 = const()[name = tensor("op_1056"), val = tensor([1, 1, 1, 1])]; + tensor var_1057 = reshape(shape = var_1056, x = var_1055)[name = tensor("op_1057")]; + tensor cross_9 = mul(x = var_1054, y = var_1057)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1060 = const()[name = tensor("op_1060"), val = tensor([1, 1, 1, 1])]; + tensor var_1061 = reshape(shape = var_1060, x = valid_mask)[name = tensor("op_1061")]; + tensor v_masked_1 = mul(x = v_9, y = var_1061)[name = tensor("v_masked_1")]; + tensor var_1063 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1063")]; + tensor var_1065_transpose_x_1 = const()[name = tensor("op_1065_transpose_x_1"), val = tensor(true)]; + tensor var_1065_transpose_y_1 = const()[name = tensor("op_1065_transpose_y_1"), val = tensor(false)]; + tensor var_1065 = matmul(transpose_x = var_1065_transpose_x_1, transpose_y = var_1065_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1065")]; + tensor new_kv_unnorm_9 = add(x = var_1063, y = var_1065)[name = tensor("new_kv_unnorm_9")]; + tensor var_1067_keep_dims_0 = const()[name = tensor("op_1067_keep_dims_0"), val = tensor(false)]; + tensor var_1067 = reduce_sum(keep_dims = var_1067_keep_dims_0, x = valid_mask)[name = tensor("op_1067")]; + tensor var_1068 = const()[name = tensor("op_1068"), val = tensor([1])]; + tensor var_1069 = reshape(shape = var_1068, x = var_1067)[name = tensor("op_1069")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1069)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_926, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1073 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1073")]; + tensor var_1074_perm_0 = const()[name = tensor("op_1074_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1074 = transpose(perm = var_1074_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_923, x = var_1074)[name = tensor("out_27")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([9, 1, 256])]; + tensor out_29 = reshape(shape = var_1078, x = out_27)[name = tensor("out_29")]; + tensor var_1080 = silu(x = input_169)[name = tensor("op_1080")]; + tensor input_171 = mul(x = var_1080, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_921, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1090 = const()[name = tensor("op_1090"), val = tensor([1, 9, 1, 256])]; + tensor var_1091 = reshape(shape = var_1090, x = xt_1)[name = tensor("op_1091")]; + tensor var_1092_perm_0 = const()[name = tensor("op_1092_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1095 = const()[name = tensor("op_1095"), val = tensor([1, 9, 256])]; + tensor var_1092 = transpose(perm = var_1092_perm_0, x = var_1091)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1095, x = var_1092)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1118 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([9, 1, 3, 256])]; + tensor var_1120 = reshape(shape = concat_1, x = var_1118)[name = tensor("op_1120")]; + tensor var_1121_axes_0 = const()[name = tensor("op_1121_axes_0"), val = tensor([0])]; + tensor var_1121 = expand_dims(axes = var_1121_axes_0, x = var_1120)[name = tensor("op_1121")]; + tensor var_1122_perm_0 = const()[name = tensor("op_1122_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1123_axes_0 = const()[name = tensor("op_1123_axes_0"), val = tensor([-2])]; + tensor var_1122 = transpose(perm = var_1122_perm_0, x = var_1121)[name = tensor("transpose_21")]; + tensor var_1123 = squeeze(axes = var_1123_axes_0, x = var_1122)[name = tensor("op_1123")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 9, 1, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1123)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 9, 1, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1123)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 9, 1, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1123)[name = tensor("v_11")]; + tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([9, 4, 64])]; + tensor var_1132 = reshape(shape = var_1131, x = q_11)[name = tensor("op_1132")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1138 = const()[name = tensor("op_1138"), val = tensor([9, 4, 64])]; + tensor var_1139 = reshape(shape = var_1138, x = k_11)[name = tensor("op_1139")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([9, 4, 64])]; + tensor var_1146 = reshape(shape = var_1145, x = v_11)[name = tensor("op_1146")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1149 = const()[name = tensor("op_1149"), val = tensor([1, 4, 9, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1132)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1149, x = q_13)[name = tensor("q_15")]; + tensor var_1151 = const()[name = tensor("op_1151"), val = tensor([1, 4, 9, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1139)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1151, x = k_13)[name = tensor("k_15")]; + tensor var_1153 = const()[name = tensor("op_1153"), val = tensor([1, 4, 9, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1146)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1153, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([2, 0, 1, 3])]; + tensor var_1161 = const()[name = tensor("op_1161"), val = tensor([9, 256])]; + tensor var_1157 = transpose(perm = var_1156, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1161, x = var_1157)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1165 = const()[name = tensor("op_1165"), val = tensor([9, 1, 256])]; + tensor attn_output_7 = reshape(shape = var_1165, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_921, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_921, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 9, 256])]; + tensor x_31 = reshape(shape = var_1185, x = xt_3)[name = tensor("x_31")]; + tensor var_1187_perm_0 = const()[name = tensor("op_1187_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([9, 1, 256])]; + tensor var_1187 = transpose(perm = var_1187_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1191, x = var_1187)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 9, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1199 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1200 = const()[name = tensor("op_1200"), val = tensor([9, 1, 4, 64])]; + tensor var_1201 = reshape(shape = var_1200, x = var_1199)[name = tensor("op_1201")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1205 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1206 = const()[name = tensor("op_1206"), val = tensor(0x1p-3)]; + tensor var_1207 = mul(x = var_1205, y = var_1206)[name = tensor("op_1207")]; + tensor var_1208 = const()[name = tensor("op_1208"), val = tensor([9, 1, 4, 64])]; + tensor var_1209 = reshape(shape = var_1208, x = var_1207)[name = tensor("op_1209")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1213 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1214 = const()[name = tensor("op_1214"), val = tensor([9, 1, 4, 64])]; + tensor var_1215 = reshape(shape = var_1214, x = var_1213)[name = tensor("op_1215")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_926, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1209)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1201)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 1])]; + tensor var_1231 = reshape(shape = var_1230, x = sqrt_s_t)[name = tensor("op_1231")]; + tensor M = real_div(x = var_1047, y = var_1231)[name = tensor("M")]; + tensor var_1233 = mul(x = qk, y = M)[name = tensor("op_1233")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1215)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1233, y = v_17)[name = tensor("inner")]; + tensor var_1235_transpose_x_0 = const()[name = tensor("op_1235_transpose_x_0"), val = tensor(false)]; + tensor var_1235_transpose_y_0 = const()[name = tensor("op_1235_transpose_y_0"), val = tensor(false)]; + tensor var_1235 = matmul(transpose_x = var_1235_transpose_x_0, transpose_y = var_1235_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1235")]; + tensor var_1236 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1236")]; + tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1, 1, 1])]; + tensor var_1238 = reshape(shape = var_1237, x = var_1236)[name = tensor("op_1238")]; + tensor cross = mul(x = var_1235, y = var_1238)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1061)[name = tensor("v_masked")]; + tensor var_1244 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1244")]; + tensor var_1246_transpose_x_1 = const()[name = tensor("op_1246_transpose_x_1"), val = tensor(true)]; + tensor var_1246_transpose_y_1 = const()[name = tensor("op_1246_transpose_y_1"), val = tensor(false)]; + tensor var_1246 = matmul(transpose_x = var_1246_transpose_x_1, transpose_y = var_1246_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1246")]; + tensor new_kv_unnorm = add(x = var_1244, y = var_1246)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1069)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_926, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1255_perm_0 = const()[name = tensor("op_1255_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1255 = transpose(perm = var_1255_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_923, x = var_1255)[name = tensor("out_33")]; + tensor var_1259 = const()[name = tensor("op_1259"), val = tensor([9, 1, 256])]; + tensor out = reshape(shape = var_1259, x = out_33)[name = tensor("out")]; + tensor var_1261 = silu(x = input_187)[name = tensor("op_1261")]; + tensor input_189 = mul(x = var_1261, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_921, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1271 = const()[name = tensor("op_1271"), val = tensor([1, 9, 1, 256])]; + tensor var_1272 = reshape(shape = var_1271, x = xt_5)[name = tensor("op_1272")]; + tensor var_1273_perm_0 = const()[name = tensor("op_1273_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1276 = const()[name = tensor("op_1276"), val = tensor([1, 9, 256])]; + tensor var_1273 = transpose(perm = var_1273_perm_0, x = var_1272)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1276, x = var_1273)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1299 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([9, 1, 3, 256])]; + tensor var_1301 = reshape(shape = concat_2, x = var_1299)[name = tensor("op_1301")]; + tensor var_1302_axes_0 = const()[name = tensor("op_1302_axes_0"), val = tensor([0])]; + tensor var_1302 = expand_dims(axes = var_1302_axes_0, x = var_1301)[name = tensor("op_1302")]; + tensor var_1303_perm_0 = const()[name = tensor("op_1303_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1304_axes_0 = const()[name = tensor("op_1304_axes_0"), val = tensor([-2])]; + tensor var_1303 = transpose(perm = var_1303_perm_0, x = var_1302)[name = tensor("transpose_8")]; + tensor var_1304 = squeeze(axes = var_1304_axes_0, x = var_1303)[name = tensor("op_1304")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 9, 1, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1304)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 9, 1, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1304)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 9, 1, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1304)[name = tensor("v_19")]; + tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([9, 4, 64])]; + tensor var_1313 = reshape(shape = var_1312, x = q_19)[name = tensor("op_1313")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1319 = const()[name = tensor("op_1319"), val = tensor([9, 4, 64])]; + tensor var_1320 = reshape(shape = var_1319, x = k_19)[name = tensor("op_1320")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([9, 4, 64])]; + tensor var_1327 = reshape(shape = var_1326, x = v_19)[name = tensor("op_1327")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1330 = const()[name = tensor("op_1330"), val = tensor([1, 4, 9, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1313)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1330, x = q_21)[name = tensor("q")]; + tensor var_1332 = const()[name = tensor("op_1332"), val = tensor([1, 4, 9, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1320)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1332, x = k_21)[name = tensor("k")]; + tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 4, 9, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1327)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1334, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([2, 0, 1, 3])]; + tensor var_1342 = const()[name = tensor("op_1342"), val = tensor([9, 256])]; + tensor var_1338 = transpose(perm = var_1337, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1342, x = var_1338)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1346 = const()[name = tensor("op_1346"), val = tensor([9, 1, 256])]; + tensor attn_output = reshape(shape = var_1346, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_921, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_921, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 9, 256])]; + tensor input = reshape(shape = var_1366, x = xt)[name = tensor("input")]; + tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([-1])]; + tensor var_1369 = reduce_l2_norm(axes = var_1368, keep_dims = var_924, x = input)[name = tensor("op_1369")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_916, beta = const_42, x = var_1369)[name = tensor("clip_5")]; + tensor var_1371 = real_div(x = input, y = clip_5)[name = tensor("op_1371")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([1, 256, 9])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1371)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = emb, y = reshape_1)[name = tensor("matmul_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 1, 8])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = matmul_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1375")]; + tensor var_1377_axis_0 = const()[name = tensor("op_1377_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1377_axis_0, values = (var_1073, nkv))[name = tensor("op_1377")]; + tensor var_1379_axis_0 = const()[name = tensor("op_1379_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1379_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1379")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/weights/weight.bin b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..df539d532d7d7a8778c91894c43572f39d8dc5eb --- /dev/null +++ b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:636026d374467f856aaee959b909ad4ec22b5101d883e3be7fe473d6b408997b +size 44404608 diff --git a/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..42c372e11d6d85d4e11b133904269e202b57e146 --- /dev/null +++ b/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a15eadb3397bff9c66d9266e9a495a92e7913dd25b985447b5efd0ebd212d182 +size 171357 diff --git a/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..df539d532d7d7a8778c91894c43572f39d8dc5eb --- /dev/null +++ b/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:636026d374467f856aaee959b909ad4ec22b5101d883e3be7fe473d6b408997b +size 44404608 diff --git a/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Manifest.json b/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..6438d33c88f42770e5903b4cf8b6b7fc7ffd893a --- /dev/null +++ b/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "15BB7841-F157-4C26-9962-7E006B1F650B": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "69383227-3EDF-4375-8C74-435BD4941612": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "15BB7841-F157-4C26-9962-7E006B1F650B" +} diff --git a/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/analytics/coremldata.bin b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a37607ed47f4304af3eb57d2df22a9ad4ac88e3e --- /dev/null +++ b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e72fc8970f10d313ad089b47fe4fcc1e390363db50c413f17b9fb0d756a8cea1 +size 243 diff --git a/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/coremldata.bin b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..fd51deca07943f75691c3895fec309e63919fa02 --- /dev/null +++ b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:230d025ab4e17869d9301adf650df4fa921472c5d5582a9f365e7d70f3f13585 +size 1301 diff --git a/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/metadata.json b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..ffa39d702bedf6118ed46a822ffe5aee24cfcb30 --- /dev/null +++ b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND CALLHOME streaming diarizer (pipeline, T=2, max_speakers=7)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 2 × 7)", + "shortDescription" : "", + "shape" : "[1, 2, 7]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 9 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 9, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 68, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 48, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 14, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 47 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 2 × 345)", + "shortDescription" : "", + "shape" : "[1, 2, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 9 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 9, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2)", + "shortDescription" : "", + "shape" : "[2]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"ch\", \"model_label\": \"CALLHOME\", \"variant\": \"pipeline\", \"chunk_size\": 2, \"step_duration_ms\": 200, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 7, \"max_nspks\": 9, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_ch_200ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/model.mil b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..801b456ed38b11b675323d063b50f9d105936618 --- /dev/null +++ b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/model.mil @@ -0,0 +1,1246 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 2, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 2, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 2, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([2, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 2, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 2, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; + tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; + tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; + tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; + tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; + tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 1, 256])]; + tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, true, true])]; + tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; + tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; + tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; + tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; + tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5))[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_257_split_sizes_0 = const()[name = tensor("op_257_split_sizes_0"), val = tensor([256, 256])]; + tensor var_257_axis_0 = const()[name = tensor("op_257_axis_0"), val = tensor(1)]; + tensor var_257_0, tensor var_257_1 = split(axis = var_257_axis_0, split_sizes = var_257_split_sizes_0, x = inputs_3)[name = tensor("op_257")]; + tensor var_259 = sigmoid(x = var_257_1)[name = tensor("op_259")]; + tensor inputs_5 = mul(x = var_257_0, y = var_259)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([2, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_290_begin_0 = const()[name = tensor("op_290_begin_0"), val = tensor([0, -1, 0])]; + tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([2, 16, 256])]; + tensor var_290_end_mask_0 = const()[name = tensor("op_290_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_290 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, x = conv_out_1)[name = tensor("op_290")]; + tensor var_292_perm_0 = const()[name = tensor("op_292_perm_0"), val = tensor([1, 0, 2])]; + tensor var_292 = transpose(perm = var_292_perm_0, x = var_290)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_292)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_315 = const()[name = tensor("op_315"), val = tensor(0x1p-1)]; + tensor var_316 = mul(x = input_39, y = var_315)[name = tensor("op_316")]; + tensor input_41 = add(x = var_316, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_345 = const()[name = tensor("op_345"), val = tensor(0x1p-1)]; + tensor var_346 = mul(x = input_51, y = var_345)[name = tensor("op_346")]; + tensor input_53 = add(x = var_346, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_360 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 2, 4, 64])]; + tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_366 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_367 = const()[name = tensor("op_367"), val = tensor(0x1p-3)]; + tensor var_368 = mul(x = var_366, y = var_367)[name = tensor("op_368")]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 2, 4, 64])]; + tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_374 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_375 = const()[name = tensor("op_375"), val = tensor([1, 2, 4, 64])]; + tensor var_376 = reshape(shape = var_375, x = var_374)[name = tensor("op_376")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_370)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_362)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([2, 1])]; + tensor var_387 = reshape(shape = var_386, x = sqrt_s_t_3)[name = tensor("op_387")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_387)[name = tensor("M_3")]; + tensor var_389 = mul(x = qk_3, y = M_3)[name = tensor("op_389")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_376)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_389, y = v_3)[name = tensor("inner_3")]; + tensor var_391_transpose_x_0 = const()[name = tensor("op_391_transpose_x_0"), val = tensor(false)]; + tensor var_391_transpose_y_0 = const()[name = tensor("op_391_transpose_y_0"), val = tensor(false)]; + tensor var_391 = matmul(transpose_x = var_391_transpose_x_0, transpose_y = var_391_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_391")]; + tensor var_392 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_392")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor([1, 1, 2, 1])]; + tensor var_394 = reshape(shape = var_393, x = var_392)[name = tensor("op_394")]; + tensor cross_3 = mul(x = var_391, y = var_394)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_397 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_397")]; + tensor var_399_transpose_x_1 = const()[name = tensor("op_399_transpose_x_1"), val = tensor(true)]; + tensor var_399_transpose_y_1 = const()[name = tensor("op_399_transpose_y_1"), val = tensor(false)]; + tensor var_399 = matmul(transpose_x = var_399_transpose_x_1, transpose_y = var_399_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_399")]; + tensor new_kv_unnorm_3 = add(x = var_397, y = var_399)[name = tensor("new_kv_unnorm_3")]; + tensor var_401 = const()[name = tensor("op_401"), val = tensor(0x1p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_401)[name = tensor("new_scale_3")]; + tensor var_403 = sqrt(x = new_scale_3)[name = tensor("op_403")]; + tensor var_404 = real_div(x = new_kv_unnorm_3, y = var_403)[name = tensor("op_404")]; + tensor var_405_perm_0 = const()[name = tensor("op_405_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_405 = transpose(perm = var_405_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_405)[name = tensor("out_9")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 2, 256])]; + tensor out_11 = reshape(shape = var_409, x = out_9)[name = tensor("out_11")]; + tensor var_411 = silu(x = input_57)[name = tensor("op_411")]; + tensor input_59 = mul(x = var_411, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_7_begin_0 = const()[name = tensor("window_7_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_7_end_0 = const()[name = tensor("window_7_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_7_end_mask_0 = const()[name = tensor("window_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_7_squeeze_mask_0 = const()[name = tensor("window_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_7 = slice_by_index(begin = window_7_begin_0, end = window_7_end_0, end_mask = window_7_end_mask_0, squeeze_mask = window_7_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_7")]; + tensor var_419_begin_0 = const()[name = tensor("op_419_begin_0"), val = tensor([0, 0, 0])]; + tensor var_419_end_0 = const()[name = tensor("op_419_end_0"), val = tensor([1, 1, 256])]; + tensor var_419_end_mask_0 = const()[name = tensor("op_419_end_mask_0"), val = tensor([true, false, true])]; + tensor var_419 = slice_by_index(begin = var_419_begin_0, end = var_419_end_0, end_mask = var_419_end_mask_0, x = x_9)[name = tensor("op_419")]; + tensor var_422_begin_0 = const()[name = tensor("op_422_begin_0"), val = tensor([0, 1, 0])]; + tensor var_422_end_0 = const()[name = tensor("op_422_end_0"), val = tensor([1, 16, 256])]; + tensor var_422_end_mask_0 = const()[name = tensor("op_422_end_mask_0"), val = tensor([true, true, true])]; + tensor var_422 = slice_by_index(begin = var_422_begin_0, end = var_422_end_0, end_mask = var_422_end_mask_0, x = window_7)[name = tensor("op_422")]; + tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; + tensor window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_422, var_419))[name = tensor("window_9")]; + tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 1, 0])]; + tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; + tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, true, true])]; + tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; + tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; + tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; + tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; + tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_9, window_11))[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_455_split_sizes_0 = const()[name = tensor("op_455_split_sizes_0"), val = tensor([256, 256])]; + tensor var_455_axis_0 = const()[name = tensor("op_455_axis_0"), val = tensor(1)]; + tensor var_455_0, tensor var_455_1 = split(axis = var_455_axis_0, split_sizes = var_455_split_sizes_0, x = inputs_13)[name = tensor("op_455")]; + tensor var_457 = sigmoid(x = var_455_1)[name = tensor("op_457")]; + tensor inputs_15 = mul(x = var_455_0, y = var_457)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([2, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_488_begin_0 = const()[name = tensor("op_488_begin_0"), val = tensor([0, -1, 0])]; + tensor var_488_end_0 = const()[name = tensor("op_488_end_0"), val = tensor([2, 16, 256])]; + tensor var_488_end_mask_0 = const()[name = tensor("op_488_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_488 = slice_by_index(begin = var_488_begin_0, end = var_488_end_0, end_mask = var_488_end_mask_0, x = conv_out_3)[name = tensor("op_488")]; + tensor var_490_perm_0 = const()[name = tensor("op_490_perm_0"), val = tensor([1, 0, 2])]; + tensor var_490 = transpose(perm = var_490_perm_0, x = var_488)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_490)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_513 = const()[name = tensor("op_513"), val = tensor(0x1p-1)]; + tensor var_514 = mul(x = input_79, y = var_513)[name = tensor("op_514")]; + tensor input_81 = add(x = var_514, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_543 = const()[name = tensor("op_543"), val = tensor(0x1p-1)]; + tensor var_544 = mul(x = input_91, y = var_543)[name = tensor("op_544")]; + tensor input_93 = add(x = var_544, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_558 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_559 = const()[name = tensor("op_559"), val = tensor([1, 2, 4, 64])]; + tensor var_560 = reshape(shape = var_559, x = var_558)[name = tensor("op_560")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_564 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_565 = const()[name = tensor("op_565"), val = tensor(0x1p-3)]; + tensor var_566 = mul(x = var_564, y = var_565)[name = tensor("op_566")]; + tensor var_567 = const()[name = tensor("op_567"), val = tensor([1, 2, 4, 64])]; + tensor var_568 = reshape(shape = var_567, x = var_566)[name = tensor("op_568")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_572 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 2, 4, 64])]; + tensor var_574 = reshape(shape = var_573, x = var_572)[name = tensor("op_574")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_568)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_560)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_584 = const()[name = tensor("op_584"), val = tensor([2, 1])]; + tensor var_585 = reshape(shape = var_584, x = sqrt_s_t_5)[name = tensor("op_585")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_585)[name = tensor("M_5")]; + tensor var_587 = mul(x = qk_5, y = M_5)[name = tensor("op_587")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_574)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_587, y = v_5)[name = tensor("inner_5")]; + tensor var_589_transpose_x_0 = const()[name = tensor("op_589_transpose_x_0"), val = tensor(false)]; + tensor var_589_transpose_y_0 = const()[name = tensor("op_589_transpose_y_0"), val = tensor(false)]; + tensor var_589 = matmul(transpose_x = var_589_transpose_x_0, transpose_y = var_589_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_589")]; + tensor var_590 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_590")]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 2, 1])]; + tensor var_592 = reshape(shape = var_591, x = var_590)[name = tensor("op_592")]; + tensor cross_5 = mul(x = var_589, y = var_592)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_595 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_595")]; + tensor var_597_transpose_x_1 = const()[name = tensor("op_597_transpose_x_1"), val = tensor(true)]; + tensor var_597_transpose_y_1 = const()[name = tensor("op_597_transpose_y_1"), val = tensor(false)]; + tensor var_597 = matmul(transpose_x = var_597_transpose_x_1, transpose_y = var_597_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_597")]; + tensor new_kv_unnorm_5 = add(x = var_595, y = var_597)[name = tensor("new_kv_unnorm_5")]; + tensor var_599 = const()[name = tensor("op_599"), val = tensor(0x1p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_599)[name = tensor("new_scale_5")]; + tensor var_601 = sqrt(x = new_scale_5)[name = tensor("op_601")]; + tensor var_602 = real_div(x = new_kv_unnorm_5, y = var_601)[name = tensor("op_602")]; + tensor var_603_perm_0 = const()[name = tensor("op_603_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_603 = transpose(perm = var_603_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_603)[name = tensor("out_15")]; + tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 2, 256])]; + tensor out_17 = reshape(shape = var_607, x = out_15)[name = tensor("out_17")]; + tensor var_609 = silu(x = input_97)[name = tensor("op_609")]; + tensor input_99 = mul(x = var_609, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; + tensor var_617_begin_0 = const()[name = tensor("op_617_begin_0"), val = tensor([0, 0, 0])]; + tensor var_617_end_0 = const()[name = tensor("op_617_end_0"), val = tensor([1, 1, 256])]; + tensor var_617_end_mask_0 = const()[name = tensor("op_617_end_mask_0"), val = tensor([true, false, true])]; + tensor var_617 = slice_by_index(begin = var_617_begin_0, end = var_617_end_0, end_mask = var_617_end_mask_0, x = x_15)[name = tensor("op_617")]; + tensor var_620_begin_0 = const()[name = tensor("op_620_begin_0"), val = tensor([0, 1, 0])]; + tensor var_620_end_0 = const()[name = tensor("op_620_end_0"), val = tensor([1, 16, 256])]; + tensor var_620_end_mask_0 = const()[name = tensor("op_620_end_mask_0"), val = tensor([true, true, true])]; + tensor var_620 = slice_by_index(begin = var_620_begin_0, end = var_620_end_0, end_mask = var_620_end_mask_0, x = window_13)[name = tensor("op_620")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_620, var_617))[name = tensor("window_15")]; + tensor var_625_begin_0 = const()[name = tensor("op_625_begin_0"), val = tensor([0, 1, 0])]; + tensor var_625_end_0 = const()[name = tensor("op_625_end_0"), val = tensor([1, 1, 256])]; + tensor var_625_end_mask_0 = const()[name = tensor("op_625_end_mask_0"), val = tensor([true, true, true])]; + tensor var_625 = slice_by_index(begin = var_625_begin_0, end = var_625_end_0, end_mask = var_625_end_mask_0, x = x_15)[name = tensor("op_625")]; + tensor var_628_begin_0 = const()[name = tensor("op_628_begin_0"), val = tensor([0, 1, 0])]; + tensor var_628_end_0 = const()[name = tensor("op_628_end_0"), val = tensor([1, 16, 256])]; + tensor var_628_end_mask_0 = const()[name = tensor("op_628_end_mask_0"), val = tensor([true, true, true])]; + tensor var_628 = slice_by_index(begin = var_628_begin_0, end = var_628_end_0, end_mask = var_628_end_mask_0, x = window_15)[name = tensor("op_628")]; + tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; + tensor window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_628, var_625))[name = tensor("window_17")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_15, window_17))[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_653_split_sizes_0 = const()[name = tensor("op_653_split_sizes_0"), val = tensor([256, 256])]; + tensor var_653_axis_0 = const()[name = tensor("op_653_axis_0"), val = tensor(1)]; + tensor var_653_0, tensor var_653_1 = split(axis = var_653_axis_0, split_sizes = var_653_split_sizes_0, x = inputs_23)[name = tensor("op_653")]; + tensor var_655 = sigmoid(x = var_653_1)[name = tensor("op_655")]; + tensor inputs_25 = mul(x = var_653_0, y = var_655)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([2, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_686_begin_0 = const()[name = tensor("op_686_begin_0"), val = tensor([0, -1, 0])]; + tensor var_686_end_0 = const()[name = tensor("op_686_end_0"), val = tensor([2, 16, 256])]; + tensor var_686_end_mask_0 = const()[name = tensor("op_686_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_686 = slice_by_index(begin = var_686_begin_0, end = var_686_end_0, end_mask = var_686_end_mask_0, x = conv_out_5)[name = tensor("op_686")]; + tensor var_688_perm_0 = const()[name = tensor("op_688_perm_0"), val = tensor([1, 0, 2])]; + tensor var_688 = transpose(perm = var_688_perm_0, x = var_686)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_688)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_711 = const()[name = tensor("op_711"), val = tensor(0x1p-1)]; + tensor var_712 = mul(x = input_119, y = var_711)[name = tensor("op_712")]; + tensor input_121 = add(x = var_712, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_741 = const()[name = tensor("op_741"), val = tensor(0x1p-1)]; + tensor var_742 = mul(x = input_131, y = var_741)[name = tensor("op_742")]; + tensor input_133 = add(x = var_742, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_756 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_757 = const()[name = tensor("op_757"), val = tensor([1, 2, 4, 64])]; + tensor var_758 = reshape(shape = var_757, x = var_756)[name = tensor("op_758")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_762 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_763 = const()[name = tensor("op_763"), val = tensor(0x1p-3)]; + tensor var_764 = mul(x = var_762, y = var_763)[name = tensor("op_764")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 2, 4, 64])]; + tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_770 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 2, 4, 64])]; + tensor var_772 = reshape(shape = var_771, x = var_770)[name = tensor("op_772")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_766)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_758)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_782 = const()[name = tensor("op_782"), val = tensor([2, 1])]; + tensor var_783 = reshape(shape = var_782, x = sqrt_s_t_7)[name = tensor("op_783")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_783)[name = tensor("M_7")]; + tensor var_785 = mul(x = qk_7, y = M_7)[name = tensor("op_785")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_772)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_785, y = v_7)[name = tensor("inner_7")]; + tensor var_787_transpose_x_0 = const()[name = tensor("op_787_transpose_x_0"), val = tensor(false)]; + tensor var_787_transpose_y_0 = const()[name = tensor("op_787_transpose_y_0"), val = tensor(false)]; + tensor var_787 = matmul(transpose_x = var_787_transpose_x_0, transpose_y = var_787_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_787")]; + tensor var_788 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_788")]; + tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 1, 2, 1])]; + tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; + tensor cross_7 = mul(x = var_787, y = var_790)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_793 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_793")]; + tensor var_795_transpose_x_1 = const()[name = tensor("op_795_transpose_x_1"), val = tensor(true)]; + tensor var_795_transpose_y_1 = const()[name = tensor("op_795_transpose_y_1"), val = tensor(false)]; + tensor var_795 = matmul(transpose_x = var_795_transpose_x_1, transpose_y = var_795_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_795")]; + tensor new_kv_unnorm_7 = add(x = var_793, y = var_795)[name = tensor("new_kv_unnorm_7")]; + tensor var_797 = const()[name = tensor("op_797"), val = tensor(0x1p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_797)[name = tensor("new_scale_7")]; + tensor var_799 = sqrt(x = new_scale_7)[name = tensor("op_799")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_799)[name = tensor("nkv_1")]; + tensor var_801_perm_0 = const()[name = tensor("op_801_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_801 = transpose(perm = var_801_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_801)[name = tensor("out_21")]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 2, 256])]; + tensor out_23 = reshape(shape = var_805, x = out_21)[name = tensor("out_23")]; + tensor var_807 = silu(x = input_137)[name = tensor("op_807")]; + tensor input_139 = mul(x = var_807, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_19_begin_0 = const()[name = tensor("window_19_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_19_end_0 = const()[name = tensor("window_19_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_19_end_mask_0 = const()[name = tensor("window_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_19_squeeze_mask_0 = const()[name = tensor("window_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_19 = slice_by_index(begin = window_19_begin_0, end = window_19_end_0, end_mask = window_19_end_mask_0, squeeze_mask = window_19_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_19")]; + tensor var_815_begin_0 = const()[name = tensor("op_815_begin_0"), val = tensor([0, 0, 0])]; + tensor var_815_end_0 = const()[name = tensor("op_815_end_0"), val = tensor([1, 1, 256])]; + tensor var_815_end_mask_0 = const()[name = tensor("op_815_end_mask_0"), val = tensor([true, false, true])]; + tensor var_815 = slice_by_index(begin = var_815_begin_0, end = var_815_end_0, end_mask = var_815_end_mask_0, x = x_21)[name = tensor("op_815")]; + tensor var_818_begin_0 = const()[name = tensor("op_818_begin_0"), val = tensor([0, 1, 0])]; + tensor var_818_end_0 = const()[name = tensor("op_818_end_0"), val = tensor([1, 16, 256])]; + tensor var_818_end_mask_0 = const()[name = tensor("op_818_end_mask_0"), val = tensor([true, true, true])]; + tensor var_818 = slice_by_index(begin = var_818_begin_0, end = var_818_end_0, end_mask = var_818_end_mask_0, x = window_19)[name = tensor("op_818")]; + tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; + tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_818, var_815))[name = tensor("window_21")]; + tensor var_823_begin_0 = const()[name = tensor("op_823_begin_0"), val = tensor([0, 1, 0])]; + tensor var_823_end_0 = const()[name = tensor("op_823_end_0"), val = tensor([1, 1, 256])]; + tensor var_823_end_mask_0 = const()[name = tensor("op_823_end_mask_0"), val = tensor([true, true, true])]; + tensor var_823 = slice_by_index(begin = var_823_begin_0, end = var_823_end_0, end_mask = var_823_end_mask_0, x = x_21)[name = tensor("op_823")]; + tensor var_826_begin_0 = const()[name = tensor("op_826_begin_0"), val = tensor([0, 1, 0])]; + tensor var_826_end_0 = const()[name = tensor("op_826_end_0"), val = tensor([1, 16, 256])]; + tensor var_826_end_mask_0 = const()[name = tensor("op_826_end_mask_0"), val = tensor([true, true, true])]; + tensor var_826 = slice_by_index(begin = var_826_begin_0, end = var_826_end_0, end_mask = var_826_end_mask_0, x = window_21)[name = tensor("op_826")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_826, var_823))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_21, window))[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_851_split_sizes_0 = const()[name = tensor("op_851_split_sizes_0"), val = tensor([256, 256])]; + tensor var_851_axis_0 = const()[name = tensor("op_851_axis_0"), val = tensor(1)]; + tensor var_851_0, tensor var_851_1 = split(axis = var_851_axis_0, split_sizes = var_851_split_sizes_0, x = inputs_33)[name = tensor("op_851")]; + tensor var_853 = sigmoid(x = var_851_1)[name = tensor("op_853")]; + tensor inputs_35 = mul(x = var_851_0, y = var_853)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([2, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_884_begin_0 = const()[name = tensor("op_884_begin_0"), val = tensor([0, -1, 0])]; + tensor var_884_end_0 = const()[name = tensor("op_884_end_0"), val = tensor([2, 16, 256])]; + tensor var_884_end_mask_0 = const()[name = tensor("op_884_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_884 = slice_by_index(begin = var_884_begin_0, end = var_884_end_0, end_mask = var_884_end_mask_0, x = conv_out_7)[name = tensor("op_884")]; + tensor var_886_perm_0 = const()[name = tensor("op_886_perm_0"), val = tensor([1, 0, 2])]; + tensor var_886 = transpose(perm = var_886_perm_0, x = var_884)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_886)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_909 = const()[name = tensor("op_909"), val = tensor(0x1p-1)]; + tensor var_910 = mul(x = input_159, y = var_909)[name = tensor("op_910")]; + tensor input_161 = add(x = var_910, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_928_begin_0 = const()[name = tensor("op_928_begin_0"), val = tensor([0, 0, 2])]; + tensor var_928_end_0 = const()[name = tensor("op_928_end_0"), val = tensor([1, 256, 20])]; + tensor var_928_end_mask_0 = const()[name = tensor("op_928_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_928_begin_0, end = var_928_end_0, end_mask = var_928_end_mask_0, x = cat)[name = tensor("op_928")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_930 = const()[name = tensor("op_930"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_931 = reduce_l2_norm(axes = var_930, keep_dims = var_29, x = input_163)[name = tensor("op_931")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_931)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_935_axis_0 = const()[name = tensor("op_935_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_935_axis_0, values = (var_206, var_404, var_602, nkv_1))[name = tensor("op_935")]; + tensor var_937_axis_0 = const()[name = tensor("op_937_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_937_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_937")]; + tensor var_939_axis_0 = const()[name = tensor("op_939_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_939_axis_0, values = (window_5, window_11, window_17, window))[name = tensor("op_939")]; + tensor var_948 = const()[name = tensor("op_948"), val = tensor(0x1.5798eep-27)]; + tensor var_953 = const()[name = tensor("op_953"), val = tensor(0x1.4f8b58p-17)]; + tensor var_955 = const()[name = tensor("op_955"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_956 = const()[name = tensor("op_956"), val = tensor(true)]; + tensor var_958 = const()[name = tensor("op_958"), val = tensor(0x1p+0)]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor(-1)]; + tensor var_968 = const()[name = tensor("op_968"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_1030_axes_0 = const()[name = tensor("op_1030_axes_0"), val = tensor([2])]; + tensor var_1030 = expand_dims(axes = var_1030_axes_0, x = emb)[name = tensor("op_1030")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 9, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1030)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_962, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1038_perm_0 = const()[name = tensor("op_1038_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1042 = const()[name = tensor("op_1042"), val = tensor([9, 2, 256])]; + tensor var_1038 = transpose(perm = var_1038_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1042, x = var_1038)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 9, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1050 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1051 = const()[name = tensor("op_1051"), val = tensor([9, 2, 4, 64])]; + tensor var_1052 = reshape(shape = var_1051, x = var_1050)[name = tensor("op_1052")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1056 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1057 = const()[name = tensor("op_1057"), val = tensor(0x1p-3)]; + tensor var_1058 = mul(x = var_1056, y = var_1057)[name = tensor("op_1058")]; + tensor var_1059 = const()[name = tensor("op_1059"), val = tensor([9, 2, 4, 64])]; + tensor var_1060 = reshape(shape = var_1059, x = var_1058)[name = tensor("op_1060")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1064 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1065 = const()[name = tensor("op_1065"), val = tensor([9, 2, 4, 64])]; + tensor var_1066 = reshape(shape = var_1065, x = var_1064)[name = tensor("op_1066")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_968, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_958, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1060)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1052)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([1, 2])]; + tensor var_1079 = reshape(shape = var_1078, x = valid_mask)[name = tensor("op_1079")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1079)[name = tensor("causal_with_valid_1")]; + tensor var_1081 = const()[name = tensor("op_1081"), val = tensor([2, 1])]; + tensor var_1082 = reshape(shape = var_1081, x = sqrt_s_t_9)[name = tensor("op_1082")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1082)[name = tensor("M_9")]; + tensor var_1084 = mul(x = qk_9, y = M_9)[name = tensor("op_1084")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1066)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1084, y = v_9)[name = tensor("inner_9")]; + tensor var_1086_transpose_x_0 = const()[name = tensor("op_1086_transpose_x_0"), val = tensor(false)]; + tensor var_1086_transpose_y_0 = const()[name = tensor("op_1086_transpose_y_0"), val = tensor(false)]; + tensor var_1086 = matmul(transpose_x = var_1086_transpose_x_0, transpose_y = var_1086_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1086")]; + tensor var_1087 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1087")]; + tensor var_1088 = const()[name = tensor("op_1088"), val = tensor([1, 1, 2, 1])]; + tensor var_1089 = reshape(shape = var_1088, x = var_1087)[name = tensor("op_1089")]; + tensor cross_9 = mul(x = var_1086, y = var_1089)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([1, 1, 2, 1])]; + tensor var_1093 = reshape(shape = var_1092, x = valid_mask)[name = tensor("op_1093")]; + tensor v_masked_1 = mul(x = v_9, y = var_1093)[name = tensor("v_masked_1")]; + tensor var_1095 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1095")]; + tensor var_1097_transpose_x_1 = const()[name = tensor("op_1097_transpose_x_1"), val = tensor(true)]; + tensor var_1097_transpose_y_1 = const()[name = tensor("op_1097_transpose_y_1"), val = tensor(false)]; + tensor var_1097 = matmul(transpose_x = var_1097_transpose_x_1, transpose_y = var_1097_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1097")]; + tensor new_kv_unnorm_9 = add(x = var_1095, y = var_1097)[name = tensor("new_kv_unnorm_9")]; + tensor var_1099_keep_dims_0 = const()[name = tensor("op_1099_keep_dims_0"), val = tensor(false)]; + tensor var_1099 = reduce_sum(keep_dims = var_1099_keep_dims_0, x = valid_mask)[name = tensor("op_1099")]; + tensor var_1100 = const()[name = tensor("op_1100"), val = tensor([1])]; + tensor var_1101 = reshape(shape = var_1100, x = var_1099)[name = tensor("op_1101")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1101)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_958, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1105 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1105")]; + tensor var_1106_perm_0 = const()[name = tensor("op_1106_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1106 = transpose(perm = var_1106_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_955, x = var_1106)[name = tensor("out_27")]; + tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([9, 2, 256])]; + tensor out_29 = reshape(shape = var_1110, x = out_27)[name = tensor("out_29")]; + tensor var_1112 = silu(x = input_169)[name = tensor("op_1112")]; + tensor input_171 = mul(x = var_1112, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_953, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1122 = const()[name = tensor("op_1122"), val = tensor([1, 9, 2, 256])]; + tensor var_1123 = reshape(shape = var_1122, x = xt_1)[name = tensor("op_1123")]; + tensor var_1124_perm_0 = const()[name = tensor("op_1124_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([2, 9, 256])]; + tensor var_1124 = transpose(perm = var_1124_perm_0, x = var_1123)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1127, x = var_1124)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1150 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([9, 2, 3, 256])]; + tensor var_1152 = reshape(shape = concat_1, x = var_1150)[name = tensor("op_1152")]; + tensor var_1153_axes_0 = const()[name = tensor("op_1153_axes_0"), val = tensor([0])]; + tensor var_1153 = expand_dims(axes = var_1153_axes_0, x = var_1152)[name = tensor("op_1153")]; + tensor var_1154_perm_0 = const()[name = tensor("op_1154_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1155_axes_0 = const()[name = tensor("op_1155_axes_0"), val = tensor([-2])]; + tensor var_1154 = transpose(perm = var_1154_perm_0, x = var_1153)[name = tensor("transpose_21")]; + tensor var_1155 = squeeze(axes = var_1155_axes_0, x = var_1154)[name = tensor("op_1155")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 9, 2, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1155)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 9, 2, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1155)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 9, 2, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1155)[name = tensor("v_11")]; + tensor var_1163 = const()[name = tensor("op_1163"), val = tensor([9, 8, 64])]; + tensor var_1164 = reshape(shape = var_1163, x = q_11)[name = tensor("op_1164")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1170 = const()[name = tensor("op_1170"), val = tensor([9, 8, 64])]; + tensor var_1171 = reshape(shape = var_1170, x = k_11)[name = tensor("op_1171")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor([9, 8, 64])]; + tensor var_1178 = reshape(shape = var_1177, x = v_11)[name = tensor("op_1178")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1181 = const()[name = tensor("op_1181"), val = tensor([2, 4, 9, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1164)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1181, x = q_13)[name = tensor("q_15")]; + tensor var_1183 = const()[name = tensor("op_1183"), val = tensor([2, 4, 9, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1171)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1183, x = k_13)[name = tensor("k_15")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([2, 4, 9, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1178)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1185, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1188 = const()[name = tensor("op_1188"), val = tensor([2, 0, 1, 3])]; + tensor var_1193 = const()[name = tensor("op_1193"), val = tensor([18, 256])]; + tensor var_1189 = transpose(perm = var_1188, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1193, x = var_1189)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([9, 2, 256])]; + tensor attn_output_7 = reshape(shape = var_1197, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_953, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_953, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([1, 2, 9, 256])]; + tensor x_31 = reshape(shape = var_1217, x = xt_3)[name = tensor("x_31")]; + tensor var_1219_perm_0 = const()[name = tensor("op_1219_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1223 = const()[name = tensor("op_1223"), val = tensor([9, 2, 256])]; + tensor var_1219 = transpose(perm = var_1219_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1223, x = var_1219)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 9, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1231 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([9, 2, 4, 64])]; + tensor var_1233 = reshape(shape = var_1232, x = var_1231)[name = tensor("op_1233")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1237 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor(0x1p-3)]; + tensor var_1239 = mul(x = var_1237, y = var_1238)[name = tensor("op_1239")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([9, 2, 4, 64])]; + tensor var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor("op_1241")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1245 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1246 = const()[name = tensor("op_1246"), val = tensor([9, 2, 4, 64])]; + tensor var_1247 = reshape(shape = var_1246, x = var_1245)[name = tensor("op_1247")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_958, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1241)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1233)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([2, 1])]; + tensor var_1263 = reshape(shape = var_1262, x = sqrt_s_t)[name = tensor("op_1263")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1263)[name = tensor("M")]; + tensor var_1265 = mul(x = qk, y = M)[name = tensor("op_1265")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1247)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1265, y = v_17)[name = tensor("inner")]; + tensor var_1267_transpose_x_0 = const()[name = tensor("op_1267_transpose_x_0"), val = tensor(false)]; + tensor var_1267_transpose_y_0 = const()[name = tensor("op_1267_transpose_y_0"), val = tensor(false)]; + tensor var_1267 = matmul(transpose_x = var_1267_transpose_x_0, transpose_y = var_1267_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1267")]; + tensor var_1268 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1268")]; + tensor var_1269 = const()[name = tensor("op_1269"), val = tensor([1, 1, 2, 1])]; + tensor var_1270 = reshape(shape = var_1269, x = var_1268)[name = tensor("op_1270")]; + tensor cross = mul(x = var_1267, y = var_1270)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1093)[name = tensor("v_masked")]; + tensor var_1276 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1276")]; + tensor var_1278_transpose_x_1 = const()[name = tensor("op_1278_transpose_x_1"), val = tensor(true)]; + tensor var_1278_transpose_y_1 = const()[name = tensor("op_1278_transpose_y_1"), val = tensor(false)]; + tensor var_1278 = matmul(transpose_x = var_1278_transpose_x_1, transpose_y = var_1278_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1278")]; + tensor new_kv_unnorm = add(x = var_1276, y = var_1278)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1101)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_958, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1287_perm_0 = const()[name = tensor("op_1287_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1287 = transpose(perm = var_1287_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_955, x = var_1287)[name = tensor("out_33")]; + tensor var_1291 = const()[name = tensor("op_1291"), val = tensor([9, 2, 256])]; + tensor out = reshape(shape = var_1291, x = out_33)[name = tensor("out")]; + tensor var_1293 = silu(x = input_187)[name = tensor("op_1293")]; + tensor input_189 = mul(x = var_1293, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_953, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1303 = const()[name = tensor("op_1303"), val = tensor([1, 9, 2, 256])]; + tensor var_1304 = reshape(shape = var_1303, x = xt_5)[name = tensor("op_1304")]; + tensor var_1305_perm_0 = const()[name = tensor("op_1305_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1308 = const()[name = tensor("op_1308"), val = tensor([2, 9, 256])]; + tensor var_1305 = transpose(perm = var_1305_perm_0, x = var_1304)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1308, x = var_1305)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1331 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([9, 2, 3, 256])]; + tensor var_1333 = reshape(shape = concat_2, x = var_1331)[name = tensor("op_1333")]; + tensor var_1334_axes_0 = const()[name = tensor("op_1334_axes_0"), val = tensor([0])]; + tensor var_1334 = expand_dims(axes = var_1334_axes_0, x = var_1333)[name = tensor("op_1334")]; + tensor var_1335_perm_0 = const()[name = tensor("op_1335_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1336_axes_0 = const()[name = tensor("op_1336_axes_0"), val = tensor([-2])]; + tensor var_1335 = transpose(perm = var_1335_perm_0, x = var_1334)[name = tensor("transpose_8")]; + tensor var_1336 = squeeze(axes = var_1336_axes_0, x = var_1335)[name = tensor("op_1336")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 9, 2, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1336)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 9, 2, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1336)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 9, 2, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1336)[name = tensor("v_19")]; + tensor var_1344 = const()[name = tensor("op_1344"), val = tensor([9, 8, 64])]; + tensor var_1345 = reshape(shape = var_1344, x = q_19)[name = tensor("op_1345")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1351 = const()[name = tensor("op_1351"), val = tensor([9, 8, 64])]; + tensor var_1352 = reshape(shape = var_1351, x = k_19)[name = tensor("op_1352")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([9, 8, 64])]; + tensor var_1359 = reshape(shape = var_1358, x = v_19)[name = tensor("op_1359")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1362 = const()[name = tensor("op_1362"), val = tensor([2, 4, 9, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1345)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1362, x = q_21)[name = tensor("q")]; + tensor var_1364 = const()[name = tensor("op_1364"), val = tensor([2, 4, 9, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1352)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1364, x = k_21)[name = tensor("k")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([2, 4, 9, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1359)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1366, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([2, 0, 1, 3])]; + tensor var_1374 = const()[name = tensor("op_1374"), val = tensor([18, 256])]; + tensor var_1370 = transpose(perm = var_1369, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1374, x = var_1370)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1378 = const()[name = tensor("op_1378"), val = tensor([9, 2, 256])]; + tensor attn_output = reshape(shape = var_1378, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_953, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_953, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([1, 2, 9, 256])]; + tensor input = reshape(shape = var_1398, x = xt)[name = tensor("input")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([-1])]; + tensor var_1401 = reduce_l2_norm(axes = var_1400, keep_dims = var_956, x = input)[name = tensor("op_1401")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_948, beta = const_42, x = var_1401)[name = tensor("clip_5")]; + tensor var_1403 = real_div(x = input, y = clip_5)[name = tensor("op_1403")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([2, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([2, 256, 9])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1403)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 2, 9])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 2, 8])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1407")]; + tensor var_1409_axis_0 = const()[name = tensor("op_1409_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1409_axis_0, values = (var_1105, nkv))[name = tensor("op_1409")]; + tensor var_1411_axis_0 = const()[name = tensor("op_1411_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1411_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1411")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/weights/weight.bin b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..2d4563c0c6923a5b0ca22ee611064a00d7389bbe --- /dev/null +++ b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef47a8e4696a7212034d4436598ba257ac25bdb82e66df1b468a8a9f4cea25d9 +size 44413824 diff --git a/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..bee7293d1930f8df17e82f398224b087ead0eed6 --- /dev/null +++ b/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f15d275fba79646bf8e6b1acd94f18b344b4ba451e802249a4814d7601f426f +size 179867 diff --git a/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..2d4563c0c6923a5b0ca22ee611064a00d7389bbe --- /dev/null +++ b/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef47a8e4696a7212034d4436598ba257ac25bdb82e66df1b468a8a9f4cea25d9 +size 44413824 diff --git a/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Manifest.json b/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..cd59efc886fa4d1cfef4032c5411dbb7569655b1 --- /dev/null +++ b/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "5C0B804E-E764-41B0-9747-F2E8A83C47B7": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "9610BCAB-14A1-4087-9959-ED1CB79286B1": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "5C0B804E-E764-41B0-9747-F2E8A83C47B7" +} diff --git a/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/analytics/coremldata.bin b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d8f91c11eb8b5980108c3d2fdd2be965c635d965 --- /dev/null +++ b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ceb496d3598e7070ae0538a32010b7e9d3ffdd68c9c569f14b6987731dc9f1ad +size 243 diff --git a/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/coremldata.bin b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..58a2886f988ba53cdc741a3c5780fe599407acd9 --- /dev/null +++ b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:595472cafa7ff618c858be86076cfe0e894ffc64361b5318c631e631bd0a0669 +size 1301 diff --git a/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/metadata.json b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..637d583fb946317262e183e23421f1a720e2a1b1 --- /dev/null +++ b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND CALLHOME streaming diarizer (pipeline, T=3, max_speakers=7)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 3 × 7)", + "shortDescription" : "", + "shape" : "[1, 3, 7]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 9 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 9, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 68, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 56, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 18, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 47 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 3 × 345)", + "shortDescription" : "", + "shape" : "[1, 3, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 9 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 9, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 3)", + "shortDescription" : "", + "shape" : "[3]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"ch\", \"model_label\": \"CALLHOME\", \"variant\": \"pipeline\", \"chunk_size\": 3, \"step_duration_ms\": 300, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 7, \"max_nspks\": 9, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_ch_300ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/model.mil b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..5eef33ebf1d0746b1ddb3ff22edac4fc03ac7b54 --- /dev/null +++ b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/model.mil @@ -0,0 +1,1286 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 3, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 3, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 3, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([3, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 3, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1.8p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 3, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; + tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; + tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; + tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; + tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; + tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; + tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; + tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; + tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; + tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; + tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; + tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; + tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; + tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 1, 256])]; + tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, true, true])]; + tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; + tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; + tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; + tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; + tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7))[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_265_split_sizes_0 = const()[name = tensor("op_265_split_sizes_0"), val = tensor([256, 256])]; + tensor var_265_axis_0 = const()[name = tensor("op_265_axis_0"), val = tensor(1)]; + tensor var_265_0, tensor var_265_1 = split(axis = var_265_axis_0, split_sizes = var_265_split_sizes_0, x = inputs_3)[name = tensor("op_265")]; + tensor var_267 = sigmoid(x = var_265_1)[name = tensor("op_267")]; + tensor inputs_5 = mul(x = var_265_0, y = var_267)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([3, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_298_begin_0 = const()[name = tensor("op_298_begin_0"), val = tensor([0, -1, 0])]; + tensor var_298_end_0 = const()[name = tensor("op_298_end_0"), val = tensor([3, 16, 256])]; + tensor var_298_end_mask_0 = const()[name = tensor("op_298_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_298 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, x = conv_out_1)[name = tensor("op_298")]; + tensor var_300_perm_0 = const()[name = tensor("op_300_perm_0"), val = tensor([1, 0, 2])]; + tensor var_300 = transpose(perm = var_300_perm_0, x = var_298)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_300)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_323 = const()[name = tensor("op_323"), val = tensor(0x1p-1)]; + tensor var_324 = mul(x = input_39, y = var_323)[name = tensor("op_324")]; + tensor input_41 = add(x = var_324, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_353 = const()[name = tensor("op_353"), val = tensor(0x1p-1)]; + tensor var_354 = mul(x = input_51, y = var_353)[name = tensor("op_354")]; + tensor input_53 = add(x = var_354, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_368 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 3, 4, 64])]; + tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_374 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_375 = const()[name = tensor("op_375"), val = tensor(0x1p-3)]; + tensor var_376 = mul(x = var_374, y = var_375)[name = tensor("op_376")]; + tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 3, 4, 64])]; + tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_382 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_383 = const()[name = tensor("op_383"), val = tensor([1, 3, 4, 64])]; + tensor var_384 = reshape(shape = var_383, x = var_382)[name = tensor("op_384")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_378)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_370)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_394 = const()[name = tensor("op_394"), val = tensor([3, 1])]; + tensor var_395 = reshape(shape = var_394, x = sqrt_s_t_3)[name = tensor("op_395")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_395)[name = tensor("M_3")]; + tensor var_397 = mul(x = qk_3, y = M_3)[name = tensor("op_397")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_384)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_397, y = v_3)[name = tensor("inner_3")]; + tensor var_399_transpose_x_0 = const()[name = tensor("op_399_transpose_x_0"), val = tensor(false)]; + tensor var_399_transpose_y_0 = const()[name = tensor("op_399_transpose_y_0"), val = tensor(false)]; + tensor var_399 = matmul(transpose_x = var_399_transpose_x_0, transpose_y = var_399_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_399")]; + tensor var_400 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_400")]; + tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 3, 1])]; + tensor var_402 = reshape(shape = var_401, x = var_400)[name = tensor("op_402")]; + tensor cross_3 = mul(x = var_399, y = var_402)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_405 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_405")]; + tensor var_407_transpose_x_1 = const()[name = tensor("op_407_transpose_x_1"), val = tensor(true)]; + tensor var_407_transpose_y_1 = const()[name = tensor("op_407_transpose_y_1"), val = tensor(false)]; + tensor var_407 = matmul(transpose_x = var_407_transpose_x_1, transpose_y = var_407_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_407")]; + tensor new_kv_unnorm_3 = add(x = var_405, y = var_407)[name = tensor("new_kv_unnorm_3")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor(0x1.8p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_409)[name = tensor("new_scale_3")]; + tensor var_411 = sqrt(x = new_scale_3)[name = tensor("op_411")]; + tensor var_412 = real_div(x = new_kv_unnorm_3, y = var_411)[name = tensor("op_412")]; + tensor var_413_perm_0 = const()[name = tensor("op_413_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_413 = transpose(perm = var_413_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_413)[name = tensor("out_9")]; + tensor var_417 = const()[name = tensor("op_417"), val = tensor([1, 3, 256])]; + tensor out_11 = reshape(shape = var_417, x = out_9)[name = tensor("out_11")]; + tensor var_419 = silu(x = input_57)[name = tensor("op_419")]; + tensor input_59 = mul(x = var_419, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; + tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 0, 0])]; + tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; + tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, false, true])]; + tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; + tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; + tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; + tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; + tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; + tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 1, 0])]; + tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 2, 256])]; + tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; + tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; + tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; + tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; + tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; + tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; + tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; + tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 2, 0])]; + tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 1, 256])]; + tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, true, true])]; + tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; + tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; + tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; + tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; + tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_11, window_13, window_15))[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_471_split_sizes_0 = const()[name = tensor("op_471_split_sizes_0"), val = tensor([256, 256])]; + tensor var_471_axis_0 = const()[name = tensor("op_471_axis_0"), val = tensor(1)]; + tensor var_471_0, tensor var_471_1 = split(axis = var_471_axis_0, split_sizes = var_471_split_sizes_0, x = inputs_13)[name = tensor("op_471")]; + tensor var_473 = sigmoid(x = var_471_1)[name = tensor("op_473")]; + tensor inputs_15 = mul(x = var_471_0, y = var_473)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([3, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_504_begin_0 = const()[name = tensor("op_504_begin_0"), val = tensor([0, -1, 0])]; + tensor var_504_end_0 = const()[name = tensor("op_504_end_0"), val = tensor([3, 16, 256])]; + tensor var_504_end_mask_0 = const()[name = tensor("op_504_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_504 = slice_by_index(begin = var_504_begin_0, end = var_504_end_0, end_mask = var_504_end_mask_0, x = conv_out_3)[name = tensor("op_504")]; + tensor var_506_perm_0 = const()[name = tensor("op_506_perm_0"), val = tensor([1, 0, 2])]; + tensor var_506 = transpose(perm = var_506_perm_0, x = var_504)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_506)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_529 = const()[name = tensor("op_529"), val = tensor(0x1p-1)]; + tensor var_530 = mul(x = input_79, y = var_529)[name = tensor("op_530")]; + tensor input_81 = add(x = var_530, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_559 = const()[name = tensor("op_559"), val = tensor(0x1p-1)]; + tensor var_560 = mul(x = input_91, y = var_559)[name = tensor("op_560")]; + tensor input_93 = add(x = var_560, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_574 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 3, 4, 64])]; + tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_580 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_581 = const()[name = tensor("op_581"), val = tensor(0x1p-3)]; + tensor var_582 = mul(x = var_580, y = var_581)[name = tensor("op_582")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 3, 4, 64])]; + tensor var_584 = reshape(shape = var_583, x = var_582)[name = tensor("op_584")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_588 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_589 = const()[name = tensor("op_589"), val = tensor([1, 3, 4, 64])]; + tensor var_590 = reshape(shape = var_589, x = var_588)[name = tensor("op_590")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_584)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_576)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_600 = const()[name = tensor("op_600"), val = tensor([3, 1])]; + tensor var_601 = reshape(shape = var_600, x = sqrt_s_t_5)[name = tensor("op_601")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_601)[name = tensor("M_5")]; + tensor var_603 = mul(x = qk_5, y = M_5)[name = tensor("op_603")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_590)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_603, y = v_5)[name = tensor("inner_5")]; + tensor var_605_transpose_x_0 = const()[name = tensor("op_605_transpose_x_0"), val = tensor(false)]; + tensor var_605_transpose_y_0 = const()[name = tensor("op_605_transpose_y_0"), val = tensor(false)]; + tensor var_605 = matmul(transpose_x = var_605_transpose_x_0, transpose_y = var_605_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_605")]; + tensor var_606 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_606")]; + tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 1, 3, 1])]; + tensor var_608 = reshape(shape = var_607, x = var_606)[name = tensor("op_608")]; + tensor cross_5 = mul(x = var_605, y = var_608)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_611 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_611")]; + tensor var_613_transpose_x_1 = const()[name = tensor("op_613_transpose_x_1"), val = tensor(true)]; + tensor var_613_transpose_y_1 = const()[name = tensor("op_613_transpose_y_1"), val = tensor(false)]; + tensor var_613 = matmul(transpose_x = var_613_transpose_x_1, transpose_y = var_613_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_613")]; + tensor new_kv_unnorm_5 = add(x = var_611, y = var_613)[name = tensor("new_kv_unnorm_5")]; + tensor var_615 = const()[name = tensor("op_615"), val = tensor(0x1.8p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_615)[name = tensor("new_scale_5")]; + tensor var_617 = sqrt(x = new_scale_5)[name = tensor("op_617")]; + tensor var_618 = real_div(x = new_kv_unnorm_5, y = var_617)[name = tensor("op_618")]; + tensor var_619_perm_0 = const()[name = tensor("op_619_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_619 = transpose(perm = var_619_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_619)[name = tensor("out_15")]; + tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 3, 256])]; + tensor out_17 = reshape(shape = var_623, x = out_15)[name = tensor("out_17")]; + tensor var_625 = silu(x = input_97)[name = tensor("op_625")]; + tensor input_99 = mul(x = var_625, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_17_begin_0 = const()[name = tensor("window_17_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_17_end_0 = const()[name = tensor("window_17_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_17_end_mask_0 = const()[name = tensor("window_17_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_17_squeeze_mask_0 = const()[name = tensor("window_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_17 = slice_by_index(begin = window_17_begin_0, end = window_17_end_0, end_mask = window_17_end_mask_0, squeeze_mask = window_17_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_17")]; + tensor var_633_begin_0 = const()[name = tensor("op_633_begin_0"), val = tensor([0, 0, 0])]; + tensor var_633_end_0 = const()[name = tensor("op_633_end_0"), val = tensor([1, 1, 256])]; + tensor var_633_end_mask_0 = const()[name = tensor("op_633_end_mask_0"), val = tensor([true, false, true])]; + tensor var_633 = slice_by_index(begin = var_633_begin_0, end = var_633_end_0, end_mask = var_633_end_mask_0, x = x_15)[name = tensor("op_633")]; + tensor var_636_begin_0 = const()[name = tensor("op_636_begin_0"), val = tensor([0, 1, 0])]; + tensor var_636_end_0 = const()[name = tensor("op_636_end_0"), val = tensor([1, 16, 256])]; + tensor var_636_end_mask_0 = const()[name = tensor("op_636_end_mask_0"), val = tensor([true, true, true])]; + tensor var_636 = slice_by_index(begin = var_636_begin_0, end = var_636_end_0, end_mask = var_636_end_mask_0, x = window_17)[name = tensor("op_636")]; + tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; + tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_636, var_633))[name = tensor("window_19")]; + tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 1, 0])]; + tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 2, 256])]; + tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, false, true])]; + tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = x_15)[name = tensor("op_641")]; + tensor var_644_begin_0 = const()[name = tensor("op_644_begin_0"), val = tensor([0, 1, 0])]; + tensor var_644_end_0 = const()[name = tensor("op_644_end_0"), val = tensor([1, 16, 256])]; + tensor var_644_end_mask_0 = const()[name = tensor("op_644_end_mask_0"), val = tensor([true, true, true])]; + tensor var_644 = slice_by_index(begin = var_644_begin_0, end = var_644_end_0, end_mask = var_644_end_mask_0, x = window_19)[name = tensor("op_644")]; + tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; + tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_644, var_641))[name = tensor("window_21")]; + tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 2, 0])]; + tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; + tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, true, true])]; + tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; + tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; + tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; + tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; + tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; + tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_19, window_21, window_23))[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_677_split_sizes_0 = const()[name = tensor("op_677_split_sizes_0"), val = tensor([256, 256])]; + tensor var_677_axis_0 = const()[name = tensor("op_677_axis_0"), val = tensor(1)]; + tensor var_677_0, tensor var_677_1 = split(axis = var_677_axis_0, split_sizes = var_677_split_sizes_0, x = inputs_23)[name = tensor("op_677")]; + tensor var_679 = sigmoid(x = var_677_1)[name = tensor("op_679")]; + tensor inputs_25 = mul(x = var_677_0, y = var_679)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([3, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_710_begin_0 = const()[name = tensor("op_710_begin_0"), val = tensor([0, -1, 0])]; + tensor var_710_end_0 = const()[name = tensor("op_710_end_0"), val = tensor([3, 16, 256])]; + tensor var_710_end_mask_0 = const()[name = tensor("op_710_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_710 = slice_by_index(begin = var_710_begin_0, end = var_710_end_0, end_mask = var_710_end_mask_0, x = conv_out_5)[name = tensor("op_710")]; + tensor var_712_perm_0 = const()[name = tensor("op_712_perm_0"), val = tensor([1, 0, 2])]; + tensor var_712 = transpose(perm = var_712_perm_0, x = var_710)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_712)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_735 = const()[name = tensor("op_735"), val = tensor(0x1p-1)]; + tensor var_736 = mul(x = input_119, y = var_735)[name = tensor("op_736")]; + tensor input_121 = add(x = var_736, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor(0x1p-1)]; + tensor var_766 = mul(x = input_131, y = var_765)[name = tensor("op_766")]; + tensor input_133 = add(x = var_766, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_780 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 3, 4, 64])]; + tensor var_782 = reshape(shape = var_781, x = var_780)[name = tensor("op_782")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_786 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_787 = const()[name = tensor("op_787"), val = tensor(0x1p-3)]; + tensor var_788 = mul(x = var_786, y = var_787)[name = tensor("op_788")]; + tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 3, 4, 64])]; + tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_794 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_795 = const()[name = tensor("op_795"), val = tensor([1, 3, 4, 64])]; + tensor var_796 = reshape(shape = var_795, x = var_794)[name = tensor("op_796")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_790)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_782)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_806 = const()[name = tensor("op_806"), val = tensor([3, 1])]; + tensor var_807 = reshape(shape = var_806, x = sqrt_s_t_7)[name = tensor("op_807")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_807)[name = tensor("M_7")]; + tensor var_809 = mul(x = qk_7, y = M_7)[name = tensor("op_809")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_796)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_809, y = v_7)[name = tensor("inner_7")]; + tensor var_811_transpose_x_0 = const()[name = tensor("op_811_transpose_x_0"), val = tensor(false)]; + tensor var_811_transpose_y_0 = const()[name = tensor("op_811_transpose_y_0"), val = tensor(false)]; + tensor var_811 = matmul(transpose_x = var_811_transpose_x_0, transpose_y = var_811_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_811")]; + tensor var_812 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_812")]; + tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 1, 3, 1])]; + tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; + tensor cross_7 = mul(x = var_811, y = var_814)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_817 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_817")]; + tensor var_819_transpose_x_1 = const()[name = tensor("op_819_transpose_x_1"), val = tensor(true)]; + tensor var_819_transpose_y_1 = const()[name = tensor("op_819_transpose_y_1"), val = tensor(false)]; + tensor var_819 = matmul(transpose_x = var_819_transpose_x_1, transpose_y = var_819_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_819")]; + tensor new_kv_unnorm_7 = add(x = var_817, y = var_819)[name = tensor("new_kv_unnorm_7")]; + tensor var_821 = const()[name = tensor("op_821"), val = tensor(0x1.8p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_821)[name = tensor("new_scale_7")]; + tensor var_823 = sqrt(x = new_scale_7)[name = tensor("op_823")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_823)[name = tensor("nkv_1")]; + tensor var_825_perm_0 = const()[name = tensor("op_825_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_825 = transpose(perm = var_825_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_825)[name = tensor("out_21")]; + tensor var_829 = const()[name = tensor("op_829"), val = tensor([1, 3, 256])]; + tensor out_23 = reshape(shape = var_829, x = out_21)[name = tensor("out_23")]; + tensor var_831 = silu(x = input_137)[name = tensor("op_831")]; + tensor input_139 = mul(x = var_831, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; + tensor var_839_begin_0 = const()[name = tensor("op_839_begin_0"), val = tensor([0, 0, 0])]; + tensor var_839_end_0 = const()[name = tensor("op_839_end_0"), val = tensor([1, 1, 256])]; + tensor var_839_end_mask_0 = const()[name = tensor("op_839_end_mask_0"), val = tensor([true, false, true])]; + tensor var_839 = slice_by_index(begin = var_839_begin_0, end = var_839_end_0, end_mask = var_839_end_mask_0, x = x_21)[name = tensor("op_839")]; + tensor var_842_begin_0 = const()[name = tensor("op_842_begin_0"), val = tensor([0, 1, 0])]; + tensor var_842_end_0 = const()[name = tensor("op_842_end_0"), val = tensor([1, 16, 256])]; + tensor var_842_end_mask_0 = const()[name = tensor("op_842_end_mask_0"), val = tensor([true, true, true])]; + tensor var_842 = slice_by_index(begin = var_842_begin_0, end = var_842_end_0, end_mask = var_842_end_mask_0, x = window_25)[name = tensor("op_842")]; + tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; + tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_842, var_839))[name = tensor("window_27")]; + tensor var_847_begin_0 = const()[name = tensor("op_847_begin_0"), val = tensor([0, 1, 0])]; + tensor var_847_end_0 = const()[name = tensor("op_847_end_0"), val = tensor([1, 2, 256])]; + tensor var_847_end_mask_0 = const()[name = tensor("op_847_end_mask_0"), val = tensor([true, false, true])]; + tensor var_847 = slice_by_index(begin = var_847_begin_0, end = var_847_end_0, end_mask = var_847_end_mask_0, x = x_21)[name = tensor("op_847")]; + tensor var_850_begin_0 = const()[name = tensor("op_850_begin_0"), val = tensor([0, 1, 0])]; + tensor var_850_end_0 = const()[name = tensor("op_850_end_0"), val = tensor([1, 16, 256])]; + tensor var_850_end_mask_0 = const()[name = tensor("op_850_end_mask_0"), val = tensor([true, true, true])]; + tensor var_850 = slice_by_index(begin = var_850_begin_0, end = var_850_end_0, end_mask = var_850_end_mask_0, x = window_27)[name = tensor("op_850")]; + tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; + tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_850, var_847))[name = tensor("window_29")]; + tensor var_855_begin_0 = const()[name = tensor("op_855_begin_0"), val = tensor([0, 2, 0])]; + tensor var_855_end_0 = const()[name = tensor("op_855_end_0"), val = tensor([1, 1, 256])]; + tensor var_855_end_mask_0 = const()[name = tensor("op_855_end_mask_0"), val = tensor([true, true, true])]; + tensor var_855 = slice_by_index(begin = var_855_begin_0, end = var_855_end_0, end_mask = var_855_end_mask_0, x = x_21)[name = tensor("op_855")]; + tensor var_858_begin_0 = const()[name = tensor("op_858_begin_0"), val = tensor([0, 1, 0])]; + tensor var_858_end_0 = const()[name = tensor("op_858_end_0"), val = tensor([1, 16, 256])]; + tensor var_858_end_mask_0 = const()[name = tensor("op_858_end_mask_0"), val = tensor([true, true, true])]; + tensor var_858 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = window_29)[name = tensor("op_858")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_858, var_855))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_27, window_29, window))[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_883_split_sizes_0 = const()[name = tensor("op_883_split_sizes_0"), val = tensor([256, 256])]; + tensor var_883_axis_0 = const()[name = tensor("op_883_axis_0"), val = tensor(1)]; + tensor var_883_0, tensor var_883_1 = split(axis = var_883_axis_0, split_sizes = var_883_split_sizes_0, x = inputs_33)[name = tensor("op_883")]; + tensor var_885 = sigmoid(x = var_883_1)[name = tensor("op_885")]; + tensor inputs_35 = mul(x = var_883_0, y = var_885)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([3, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_916_begin_0 = const()[name = tensor("op_916_begin_0"), val = tensor([0, -1, 0])]; + tensor var_916_end_0 = const()[name = tensor("op_916_end_0"), val = tensor([3, 16, 256])]; + tensor var_916_end_mask_0 = const()[name = tensor("op_916_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_916 = slice_by_index(begin = var_916_begin_0, end = var_916_end_0, end_mask = var_916_end_mask_0, x = conv_out_7)[name = tensor("op_916")]; + tensor var_918_perm_0 = const()[name = tensor("op_918_perm_0"), val = tensor([1, 0, 2])]; + tensor var_918 = transpose(perm = var_918_perm_0, x = var_916)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_918)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_941 = const()[name = tensor("op_941"), val = tensor(0x1p-1)]; + tensor var_942 = mul(x = input_159, y = var_941)[name = tensor("op_942")]; + tensor input_161 = add(x = var_942, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_960_begin_0 = const()[name = tensor("op_960_begin_0"), val = tensor([0, 0, 3])]; + tensor var_960_end_0 = const()[name = tensor("op_960_end_0"), val = tensor([1, 256, 21])]; + tensor var_960_end_mask_0 = const()[name = tensor("op_960_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_960_begin_0, end = var_960_end_0, end_mask = var_960_end_mask_0, x = cat)[name = tensor("op_960")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_963 = reduce_l2_norm(axes = var_962, keep_dims = var_29, x = input_163)[name = tensor("op_963")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_963)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_967_axis_0, values = (var_206, var_412, var_618, nkv_1))[name = tensor("op_967")]; + tensor var_969_axis_0 = const()[name = tensor("op_969_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_969_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_969")]; + tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_971_axis_0, values = (window_7, window_15, window_23, window))[name = tensor("op_971")]; + tensor var_980 = const()[name = tensor("op_980"), val = tensor(0x1.5798eep-27)]; + tensor var_985 = const()[name = tensor("op_985"), val = tensor(0x1.4f8b58p-17)]; + tensor var_987 = const()[name = tensor("op_987"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_988 = const()[name = tensor("op_988"), val = tensor(true)]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor(0x1p+0)]; + tensor var_994 = const()[name = tensor("op_994"), val = tensor(-1)]; + tensor var_1000 = const()[name = tensor("op_1000"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_1062_axes_0 = const()[name = tensor("op_1062_axes_0"), val = tensor([2])]; + tensor var_1062 = expand_dims(axes = var_1062_axes_0, x = emb)[name = tensor("op_1062")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 9, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1062)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_994, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1070_perm_0 = const()[name = tensor("op_1070_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1074 = const()[name = tensor("op_1074"), val = tensor([9, 3, 256])]; + tensor var_1070 = transpose(perm = var_1070_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1074, x = var_1070)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 9, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1082 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1083 = const()[name = tensor("op_1083"), val = tensor([9, 3, 4, 64])]; + tensor var_1084 = reshape(shape = var_1083, x = var_1082)[name = tensor("op_1084")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1088 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1089 = const()[name = tensor("op_1089"), val = tensor(0x1p-3)]; + tensor var_1090 = mul(x = var_1088, y = var_1089)[name = tensor("op_1090")]; + tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([9, 3, 4, 64])]; + tensor var_1092 = reshape(shape = var_1091, x = var_1090)[name = tensor("op_1092")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1096 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([9, 3, 4, 64])]; + tensor var_1098 = reshape(shape = var_1097, x = var_1096)[name = tensor("op_1098")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_1000, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_990, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1092)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1084)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([1, 3])]; + tensor var_1111 = reshape(shape = var_1110, x = valid_mask)[name = tensor("op_1111")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1111)[name = tensor("causal_with_valid_1")]; + tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([3, 1])]; + tensor var_1114 = reshape(shape = var_1113, x = sqrt_s_t_9)[name = tensor("op_1114")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1114)[name = tensor("M_9")]; + tensor var_1116 = mul(x = qk_9, y = M_9)[name = tensor("op_1116")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1098)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1116, y = v_9)[name = tensor("inner_9")]; + tensor var_1118_transpose_x_0 = const()[name = tensor("op_1118_transpose_x_0"), val = tensor(false)]; + tensor var_1118_transpose_y_0 = const()[name = tensor("op_1118_transpose_y_0"), val = tensor(false)]; + tensor var_1118 = matmul(transpose_x = var_1118_transpose_x_0, transpose_y = var_1118_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1118")]; + tensor var_1119 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1119")]; + tensor var_1120 = const()[name = tensor("op_1120"), val = tensor([1, 1, 3, 1])]; + tensor var_1121 = reshape(shape = var_1120, x = var_1119)[name = tensor("op_1121")]; + tensor cross_9 = mul(x = var_1118, y = var_1121)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1124 = const()[name = tensor("op_1124"), val = tensor([1, 1, 3, 1])]; + tensor var_1125 = reshape(shape = var_1124, x = valid_mask)[name = tensor("op_1125")]; + tensor v_masked_1 = mul(x = v_9, y = var_1125)[name = tensor("v_masked_1")]; + tensor var_1127 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1127")]; + tensor var_1129_transpose_x_1 = const()[name = tensor("op_1129_transpose_x_1"), val = tensor(true)]; + tensor var_1129_transpose_y_1 = const()[name = tensor("op_1129_transpose_y_1"), val = tensor(false)]; + tensor var_1129 = matmul(transpose_x = var_1129_transpose_x_1, transpose_y = var_1129_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1129")]; + tensor new_kv_unnorm_9 = add(x = var_1127, y = var_1129)[name = tensor("new_kv_unnorm_9")]; + tensor var_1131_keep_dims_0 = const()[name = tensor("op_1131_keep_dims_0"), val = tensor(false)]; + tensor var_1131 = reduce_sum(keep_dims = var_1131_keep_dims_0, x = valid_mask)[name = tensor("op_1131")]; + tensor var_1132 = const()[name = tensor("op_1132"), val = tensor([1])]; + tensor var_1133 = reshape(shape = var_1132, x = var_1131)[name = tensor("op_1133")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1133)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_990, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1137 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1137")]; + tensor var_1138_perm_0 = const()[name = tensor("op_1138_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1138 = transpose(perm = var_1138_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_987, x = var_1138)[name = tensor("out_27")]; + tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([9, 3, 256])]; + tensor out_29 = reshape(shape = var_1142, x = out_27)[name = tensor("out_29")]; + tensor var_1144 = silu(x = input_169)[name = tensor("op_1144")]; + tensor input_171 = mul(x = var_1144, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_985, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 9, 3, 256])]; + tensor var_1155 = reshape(shape = var_1154, x = xt_1)[name = tensor("op_1155")]; + tensor var_1156_perm_0 = const()[name = tensor("op_1156_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1159 = const()[name = tensor("op_1159"), val = tensor([3, 9, 256])]; + tensor var_1156 = transpose(perm = var_1156_perm_0, x = var_1155)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1159, x = var_1156)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1182 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([9, 3, 3, 256])]; + tensor var_1184 = reshape(shape = concat_1, x = var_1182)[name = tensor("op_1184")]; + tensor var_1185_axes_0 = const()[name = tensor("op_1185_axes_0"), val = tensor([0])]; + tensor var_1185 = expand_dims(axes = var_1185_axes_0, x = var_1184)[name = tensor("op_1185")]; + tensor var_1186_perm_0 = const()[name = tensor("op_1186_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1187_axes_0 = const()[name = tensor("op_1187_axes_0"), val = tensor([-2])]; + tensor var_1186 = transpose(perm = var_1186_perm_0, x = var_1185)[name = tensor("transpose_21")]; + tensor var_1187 = squeeze(axes = var_1187_axes_0, x = var_1186)[name = tensor("op_1187")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 9, 3, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1187)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 9, 3, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1187)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 9, 3, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1187)[name = tensor("v_11")]; + tensor var_1195 = const()[name = tensor("op_1195"), val = tensor([9, 12, 64])]; + tensor var_1196 = reshape(shape = var_1195, x = q_11)[name = tensor("op_1196")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1202 = const()[name = tensor("op_1202"), val = tensor([9, 12, 64])]; + tensor var_1203 = reshape(shape = var_1202, x = k_11)[name = tensor("op_1203")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1209 = const()[name = tensor("op_1209"), val = tensor([9, 12, 64])]; + tensor var_1210 = reshape(shape = var_1209, x = v_11)[name = tensor("op_1210")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1213 = const()[name = tensor("op_1213"), val = tensor([3, 4, 9, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1196)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1213, x = q_13)[name = tensor("q_15")]; + tensor var_1215 = const()[name = tensor("op_1215"), val = tensor([3, 4, 9, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1203)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1215, x = k_13)[name = tensor("k_15")]; + tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([3, 4, 9, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1210)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1217, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1220 = const()[name = tensor("op_1220"), val = tensor([2, 0, 1, 3])]; + tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([27, 256])]; + tensor var_1221 = transpose(perm = var_1220, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1225, x = var_1221)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1229 = const()[name = tensor("op_1229"), val = tensor([9, 3, 256])]; + tensor attn_output_7 = reshape(shape = var_1229, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_985, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_985, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([1, 3, 9, 256])]; + tensor x_31 = reshape(shape = var_1249, x = xt_3)[name = tensor("x_31")]; + tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1255 = const()[name = tensor("op_1255"), val = tensor([9, 3, 256])]; + tensor var_1251 = transpose(perm = var_1251_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1255, x = var_1251)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 9, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1263 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1264 = const()[name = tensor("op_1264"), val = tensor([9, 3, 4, 64])]; + tensor var_1265 = reshape(shape = var_1264, x = var_1263)[name = tensor("op_1265")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1269 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1270 = const()[name = tensor("op_1270"), val = tensor(0x1p-3)]; + tensor var_1271 = mul(x = var_1269, y = var_1270)[name = tensor("op_1271")]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([9, 3, 4, 64])]; + tensor var_1273 = reshape(shape = var_1272, x = var_1271)[name = tensor("op_1273")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1277 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([9, 3, 4, 64])]; + tensor var_1279 = reshape(shape = var_1278, x = var_1277)[name = tensor("op_1279")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_990, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1273)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1265)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([3, 1])]; + tensor var_1295 = reshape(shape = var_1294, x = sqrt_s_t)[name = tensor("op_1295")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1295)[name = tensor("M")]; + tensor var_1297 = mul(x = qk, y = M)[name = tensor("op_1297")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1279)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1297, y = v_17)[name = tensor("inner")]; + tensor var_1299_transpose_x_0 = const()[name = tensor("op_1299_transpose_x_0"), val = tensor(false)]; + tensor var_1299_transpose_y_0 = const()[name = tensor("op_1299_transpose_y_0"), val = tensor(false)]; + tensor var_1299 = matmul(transpose_x = var_1299_transpose_x_0, transpose_y = var_1299_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1299")]; + tensor var_1300 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1300")]; + tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([1, 1, 3, 1])]; + tensor var_1302 = reshape(shape = var_1301, x = var_1300)[name = tensor("op_1302")]; + tensor cross = mul(x = var_1299, y = var_1302)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1125)[name = tensor("v_masked")]; + tensor var_1308 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1308")]; + tensor var_1310_transpose_x_1 = const()[name = tensor("op_1310_transpose_x_1"), val = tensor(true)]; + tensor var_1310_transpose_y_1 = const()[name = tensor("op_1310_transpose_y_1"), val = tensor(false)]; + tensor var_1310 = matmul(transpose_x = var_1310_transpose_x_1, transpose_y = var_1310_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1310")]; + tensor new_kv_unnorm = add(x = var_1308, y = var_1310)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1133)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_990, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1319_perm_0 = const()[name = tensor("op_1319_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1319 = transpose(perm = var_1319_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_987, x = var_1319)[name = tensor("out_33")]; + tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([9, 3, 256])]; + tensor out = reshape(shape = var_1323, x = out_33)[name = tensor("out")]; + tensor var_1325 = silu(x = input_187)[name = tensor("op_1325")]; + tensor input_189 = mul(x = var_1325, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_985, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor([1, 9, 3, 256])]; + tensor var_1336 = reshape(shape = var_1335, x = xt_5)[name = tensor("op_1336")]; + tensor var_1337_perm_0 = const()[name = tensor("op_1337_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1340 = const()[name = tensor("op_1340"), val = tensor([3, 9, 256])]; + tensor var_1337 = transpose(perm = var_1337_perm_0, x = var_1336)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1340, x = var_1337)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1363 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([9, 3, 3, 256])]; + tensor var_1365 = reshape(shape = concat_2, x = var_1363)[name = tensor("op_1365")]; + tensor var_1366_axes_0 = const()[name = tensor("op_1366_axes_0"), val = tensor([0])]; + tensor var_1366 = expand_dims(axes = var_1366_axes_0, x = var_1365)[name = tensor("op_1366")]; + tensor var_1367_perm_0 = const()[name = tensor("op_1367_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1368_axes_0 = const()[name = tensor("op_1368_axes_0"), val = tensor([-2])]; + tensor var_1367 = transpose(perm = var_1367_perm_0, x = var_1366)[name = tensor("transpose_8")]; + tensor var_1368 = squeeze(axes = var_1368_axes_0, x = var_1367)[name = tensor("op_1368")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 9, 3, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1368)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 9, 3, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1368)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 9, 3, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1368)[name = tensor("v_19")]; + tensor var_1376 = const()[name = tensor("op_1376"), val = tensor([9, 12, 64])]; + tensor var_1377 = reshape(shape = var_1376, x = q_19)[name = tensor("op_1377")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([9, 12, 64])]; + tensor var_1384 = reshape(shape = var_1383, x = k_19)[name = tensor("op_1384")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([9, 12, 64])]; + tensor var_1391 = reshape(shape = var_1390, x = v_19)[name = tensor("op_1391")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1394 = const()[name = tensor("op_1394"), val = tensor([3, 4, 9, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1377)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1394, x = q_21)[name = tensor("q")]; + tensor var_1396 = const()[name = tensor("op_1396"), val = tensor([3, 4, 9, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1384)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1396, x = k_21)[name = tensor("k")]; + tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([3, 4, 9, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1391)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1398, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1401 = const()[name = tensor("op_1401"), val = tensor([2, 0, 1, 3])]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([27, 256])]; + tensor var_1402 = transpose(perm = var_1401, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1406, x = var_1402)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1410 = const()[name = tensor("op_1410"), val = tensor([9, 3, 256])]; + tensor attn_output = reshape(shape = var_1410, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_985, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_985, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([1, 3, 9, 256])]; + tensor input = reshape(shape = var_1430, x = xt)[name = tensor("input")]; + tensor var_1432 = const()[name = tensor("op_1432"), val = tensor([-1])]; + tensor var_1433 = reduce_l2_norm(axes = var_1432, keep_dims = var_988, x = input)[name = tensor("op_1433")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_980, beta = const_42, x = var_1433)[name = tensor("clip_5")]; + tensor var_1435 = real_div(x = input, y = clip_5)[name = tensor("op_1435")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([3, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([3, 256, 9])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1435)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 3, 9])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 3, 8])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1439")]; + tensor var_1441_axis_0 = const()[name = tensor("op_1441_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1441_axis_0, values = (var_1137, nkv))[name = tensor("op_1441")]; + tensor var_1443_axis_0 = const()[name = tensor("op_1443_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1443_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1443")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/weights/weight.bin b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..cdd4190d1046177e6ae6571103be6614a06e28b0 --- /dev/null +++ b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa68ac04a60f86f48e6a7f316418468458f69e5515eee8946fd3c260c8dcfd30 +size 44423040 diff --git a/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..09eaf79a16883ad1d32f03778b55635fbbffe830 --- /dev/null +++ b/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e633b39d71e29671336855550a3677233cf59cb49fe3e97a9c94677d5acb3f1 +size 185460 diff --git a/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..cdd4190d1046177e6ae6571103be6614a06e28b0 --- /dev/null +++ b/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa68ac04a60f86f48e6a7f316418468458f69e5515eee8946fd3c260c8dcfd30 +size 44423040 diff --git a/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Manifest.json b/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..250225980694a093a91bfcf274f8ddeff9d47549 --- /dev/null +++ b/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "539B39A7-7E21-40EC-852C-90E40A6CCF50": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "743E81BC-0B9E-4702-AC8D-3B134F65C5D1": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "539B39A7-7E21-40EC-852C-90E40A6CCF50" +} diff --git a/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/analytics/coremldata.bin b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..8ef1ae6589fe15764e0c192933da3263e5f868b8 --- /dev/null +++ b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:030a2bc7f91827f6efdac515150972512aa462ecb691365f25d97480cc537e7b +size 243 diff --git a/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/coremldata.bin b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a82c5d753734fd18061bb4de2c2b1b5334ef3f57 --- /dev/null +++ b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8fe19dbc10aceae5862d5502e24f0c7c00f356c66d3828aae97b2c41383a71b +size 1301 diff --git a/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/metadata.json b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..1479f59b57daba4565d10a86496002152bdb5208 --- /dev/null +++ b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND CALLHOME streaming diarizer (pipeline, T=4, max_speakers=7)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 4 × 7)", + "shortDescription" : "", + "shape" : "[1, 4, 7]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 9 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 9, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 68, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 64, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 22, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 47 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 4 × 345)", + "shortDescription" : "", + "shape" : "[1, 4, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 9 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 9, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4)", + "shortDescription" : "", + "shape" : "[4]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"ch\", \"model_label\": \"CALLHOME\", \"variant\": \"pipeline\", \"chunk_size\": 4, \"step_duration_ms\": 400, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 7, \"max_nspks\": 9, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_ch_400ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/model.mil b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..8e884a60ea34232c188dc4cdb669760172249293 --- /dev/null +++ b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/model.mil @@ -0,0 +1,1326 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982080)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983168)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336512)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337600)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338688)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339776)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340864)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345024)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393664)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394752)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443392)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444480)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445568)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446656)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708864)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709952)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972160)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973248)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235456)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236544)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498752)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499840)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762048)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763136)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764224)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766336)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290688)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307136)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308224)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309312)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310400)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311488)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312576)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574784)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575872)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576960)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581120)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629760)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630848)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679488)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680576)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681664)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682752)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683840)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688000)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736640)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737728)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786368)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787456)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788544)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789632)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051840)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052928)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315136)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316224)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578432)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579520)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841728)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842816)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105024)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106112)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107200)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109312)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633664)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650112)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651200)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652288)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653376)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654464)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655552)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917760)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918848)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919936)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924096)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972736)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973824)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022464)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023552)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024640)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025728)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026816)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030976)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079616)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080704)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129344)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130432)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131520)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132608)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394816)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395904)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658112)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659200)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921408)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922496)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184704)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185792)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448000)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449088)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450176)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452288)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976640)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993088)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994176)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995264)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996352)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997440)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998528)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260736)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261824)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262912)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267072)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315712)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316800)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365440)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366528)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367616)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368704)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369792)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373952)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422592)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423680)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472320)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473408)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474496)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475584)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737792)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738880)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001088)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002176)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264384)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265472)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527680)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528768)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790976)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792064)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793152)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795264)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319616)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336064)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337152)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338240)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339328)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340416)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341504)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603712)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604800)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605888)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610048)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658688)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659776)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708416)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709504)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710592)))]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710720)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711808)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236160)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237248)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499456)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500544)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762752)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763840)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026048)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027136)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289344)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290432)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552640)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553728)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554816)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555904)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818112)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821248)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607744)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608832)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609920)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618176)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715392)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716480)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813696)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814784)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815872)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816960)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079168)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080256)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342464)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343552)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605760)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606848)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869056)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870144)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132352)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133440)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134528)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135616)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397824)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400960)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187456)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188544)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189632)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197888)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295104)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296192)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393408)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394496)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 4, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 4, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 4, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([4, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 4, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 4, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; + tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; + tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; + tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; + tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; + tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; + tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; + tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; + tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; + tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; + tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; + tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; + tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; + tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 3, 256])]; + tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, false, true])]; + tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; + tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; + tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; + tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; + tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; + tensor var_245_begin_0 = const()[name = tensor("op_245_begin_0"), val = tensor([0, 3, 0])]; + tensor var_245_end_0 = const()[name = tensor("op_245_end_0"), val = tensor([1, 1, 256])]; + tensor var_245_end_mask_0 = const()[name = tensor("op_245_end_mask_0"), val = tensor([true, true, true])]; + tensor var_245 = slice_by_index(begin = var_245_begin_0, end = var_245_end_0, end_mask = var_245_end_mask_0, x = x_3)[name = tensor("op_245")]; + tensor var_248_begin_0 = const()[name = tensor("op_248_begin_0"), val = tensor([0, 1, 0])]; + tensor var_248_end_0 = const()[name = tensor("op_248_end_0"), val = tensor([1, 16, 256])]; + tensor var_248_end_mask_0 = const()[name = tensor("op_248_end_mask_0"), val = tensor([true, true, true])]; + tensor var_248 = slice_by_index(begin = var_248_begin_0, end = var_248_end_0, end_mask = var_248_end_mask_0, x = window_7)[name = tensor("op_248")]; + tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; + tensor window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_248, var_245))[name = tensor("window_9")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9))[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_273_split_sizes_0 = const()[name = tensor("op_273_split_sizes_0"), val = tensor([256, 256])]; + tensor var_273_axis_0 = const()[name = tensor("op_273_axis_0"), val = tensor(1)]; + tensor var_273_0, tensor var_273_1 = split(axis = var_273_axis_0, split_sizes = var_273_split_sizes_0, x = inputs_3)[name = tensor("op_273")]; + tensor var_275 = sigmoid(x = var_273_1)[name = tensor("op_275")]; + tensor inputs_5 = mul(x = var_273_0, y = var_275)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([4, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([0, -1, 0])]; + tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([4, 16, 256])]; + tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_306 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, x = conv_out_1)[name = tensor("op_306")]; + tensor var_308_perm_0 = const()[name = tensor("op_308_perm_0"), val = tensor([1, 0, 2])]; + tensor var_308 = transpose(perm = var_308_perm_0, x = var_306)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_308)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_331 = const()[name = tensor("op_331"), val = tensor(0x1p-1)]; + tensor var_332 = mul(x = input_39, y = var_331)[name = tensor("op_332")]; + tensor input_41 = add(x = var_332, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor(0x1p-1)]; + tensor var_362 = mul(x = input_51, y = var_361)[name = tensor("op_362")]; + tensor input_53 = add(x = var_362, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_376 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 4, 4, 64])]; + tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_382 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_383 = const()[name = tensor("op_383"), val = tensor(0x1p-3)]; + tensor var_384 = mul(x = var_382, y = var_383)[name = tensor("op_384")]; + tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 4, 4, 64])]; + tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_390 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 4, 4, 64])]; + tensor var_392 = reshape(shape = var_391, x = var_390)[name = tensor("op_392")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_386)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_378)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_402 = const()[name = tensor("op_402"), val = tensor([4, 1])]; + tensor var_403 = reshape(shape = var_402, x = sqrt_s_t_3)[name = tensor("op_403")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_403)[name = tensor("M_3")]; + tensor var_405 = mul(x = qk_3, y = M_3)[name = tensor("op_405")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_392)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_405, y = v_3)[name = tensor("inner_3")]; + tensor var_407_transpose_x_0 = const()[name = tensor("op_407_transpose_x_0"), val = tensor(false)]; + tensor var_407_transpose_y_0 = const()[name = tensor("op_407_transpose_y_0"), val = tensor(false)]; + tensor var_407 = matmul(transpose_x = var_407_transpose_x_0, transpose_y = var_407_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_407")]; + tensor var_408 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_408")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1, 4, 1])]; + tensor var_410 = reshape(shape = var_409, x = var_408)[name = tensor("op_410")]; + tensor cross_3 = mul(x = var_407, y = var_410)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_413 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_413")]; + tensor var_415_transpose_x_1 = const()[name = tensor("op_415_transpose_x_1"), val = tensor(true)]; + tensor var_415_transpose_y_1 = const()[name = tensor("op_415_transpose_y_1"), val = tensor(false)]; + tensor var_415 = matmul(transpose_x = var_415_transpose_x_1, transpose_y = var_415_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_415")]; + tensor new_kv_unnorm_3 = add(x = var_413, y = var_415)[name = tensor("new_kv_unnorm_3")]; + tensor var_417 = const()[name = tensor("op_417"), val = tensor(0x1p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_417)[name = tensor("new_scale_3")]; + tensor var_419 = sqrt(x = new_scale_3)[name = tensor("op_419")]; + tensor var_420 = real_div(x = new_kv_unnorm_3, y = var_419)[name = tensor("op_420")]; + tensor var_421_perm_0 = const()[name = tensor("op_421_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_421 = transpose(perm = var_421_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_421)[name = tensor("out_9")]; + tensor var_425 = const()[name = tensor("op_425"), val = tensor([1, 4, 256])]; + tensor out_11 = reshape(shape = var_425, x = out_9)[name = tensor("out_11")]; + tensor var_427 = silu(x = input_57)[name = tensor("op_427")]; + tensor input_59 = mul(x = var_427, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_11_begin_0 = const()[name = tensor("window_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_11_end_0 = const()[name = tensor("window_11_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_11_end_mask_0 = const()[name = tensor("window_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_11_squeeze_mask_0 = const()[name = tensor("window_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_11 = slice_by_index(begin = window_11_begin_0, end = window_11_end_0, end_mask = window_11_end_mask_0, squeeze_mask = window_11_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_11")]; + tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 0, 0])]; + tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 1, 256])]; + tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; + tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; + tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; + tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; + tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; + tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; + tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; + tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 1, 0])]; + tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 2, 256])]; + tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, false, true])]; + tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; + tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; + tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; + tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; + tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; + tensor var_451_begin_0 = const()[name = tensor("op_451_begin_0"), val = tensor([0, 2, 0])]; + tensor var_451_end_0 = const()[name = tensor("op_451_end_0"), val = tensor([1, 3, 256])]; + tensor var_451_end_mask_0 = const()[name = tensor("op_451_end_mask_0"), val = tensor([true, false, true])]; + tensor var_451 = slice_by_index(begin = var_451_begin_0, end = var_451_end_0, end_mask = var_451_end_mask_0, x = x_9)[name = tensor("op_451")]; + tensor var_454_begin_0 = const()[name = tensor("op_454_begin_0"), val = tensor([0, 1, 0])]; + tensor var_454_end_0 = const()[name = tensor("op_454_end_0"), val = tensor([1, 16, 256])]; + tensor var_454_end_mask_0 = const()[name = tensor("op_454_end_mask_0"), val = tensor([true, true, true])]; + tensor var_454 = slice_by_index(begin = var_454_begin_0, end = var_454_end_0, end_mask = var_454_end_mask_0, x = window_15)[name = tensor("op_454")]; + tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; + tensor window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_454, var_451))[name = tensor("window_17")]; + tensor var_459_begin_0 = const()[name = tensor("op_459_begin_0"), val = tensor([0, 3, 0])]; + tensor var_459_end_0 = const()[name = tensor("op_459_end_0"), val = tensor([1, 1, 256])]; + tensor var_459_end_mask_0 = const()[name = tensor("op_459_end_mask_0"), val = tensor([true, true, true])]; + tensor var_459 = slice_by_index(begin = var_459_begin_0, end = var_459_end_0, end_mask = var_459_end_mask_0, x = x_9)[name = tensor("op_459")]; + tensor var_462_begin_0 = const()[name = tensor("op_462_begin_0"), val = tensor([0, 1, 0])]; + tensor var_462_end_0 = const()[name = tensor("op_462_end_0"), val = tensor([1, 16, 256])]; + tensor var_462_end_mask_0 = const()[name = tensor("op_462_end_mask_0"), val = tensor([true, true, true])]; + tensor var_462 = slice_by_index(begin = var_462_begin_0, end = var_462_end_0, end_mask = var_462_end_mask_0, x = window_17)[name = tensor("op_462")]; + tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; + tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_462, var_459))[name = tensor("window_19")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_13, window_15, window_17, window_19))[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_487_split_sizes_0 = const()[name = tensor("op_487_split_sizes_0"), val = tensor([256, 256])]; + tensor var_487_axis_0 = const()[name = tensor("op_487_axis_0"), val = tensor(1)]; + tensor var_487_0, tensor var_487_1 = split(axis = var_487_axis_0, split_sizes = var_487_split_sizes_0, x = inputs_13)[name = tensor("op_487")]; + tensor var_489 = sigmoid(x = var_487_1)[name = tensor("op_489")]; + tensor inputs_15 = mul(x = var_487_0, y = var_489)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([4, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_520_begin_0 = const()[name = tensor("op_520_begin_0"), val = tensor([0, -1, 0])]; + tensor var_520_end_0 = const()[name = tensor("op_520_end_0"), val = tensor([4, 16, 256])]; + tensor var_520_end_mask_0 = const()[name = tensor("op_520_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_520 = slice_by_index(begin = var_520_begin_0, end = var_520_end_0, end_mask = var_520_end_mask_0, x = conv_out_3)[name = tensor("op_520")]; + tensor var_522_perm_0 = const()[name = tensor("op_522_perm_0"), val = tensor([1, 0, 2])]; + tensor var_522 = transpose(perm = var_522_perm_0, x = var_520)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_522)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_545 = const()[name = tensor("op_545"), val = tensor(0x1p-1)]; + tensor var_546 = mul(x = input_79, y = var_545)[name = tensor("op_546")]; + tensor input_81 = add(x = var_546, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-1)]; + tensor var_576 = mul(x = input_91, y = var_575)[name = tensor("op_576")]; + tensor input_93 = add(x = var_576, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_590 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 4, 4, 64])]; + tensor var_592 = reshape(shape = var_591, x = var_590)[name = tensor("op_592")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_596 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_597 = const()[name = tensor("op_597"), val = tensor(0x1p-3)]; + tensor var_598 = mul(x = var_596, y = var_597)[name = tensor("op_598")]; + tensor var_599 = const()[name = tensor("op_599"), val = tensor([1, 4, 4, 64])]; + tensor var_600 = reshape(shape = var_599, x = var_598)[name = tensor("op_600")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_604 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_605 = const()[name = tensor("op_605"), val = tensor([1, 4, 4, 64])]; + tensor var_606 = reshape(shape = var_605, x = var_604)[name = tensor("op_606")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_600)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_592)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_616 = const()[name = tensor("op_616"), val = tensor([4, 1])]; + tensor var_617 = reshape(shape = var_616, x = sqrt_s_t_5)[name = tensor("op_617")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_617)[name = tensor("M_5")]; + tensor var_619 = mul(x = qk_5, y = M_5)[name = tensor("op_619")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_606)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_619, y = v_5)[name = tensor("inner_5")]; + tensor var_621_transpose_x_0 = const()[name = tensor("op_621_transpose_x_0"), val = tensor(false)]; + tensor var_621_transpose_y_0 = const()[name = tensor("op_621_transpose_y_0"), val = tensor(false)]; + tensor var_621 = matmul(transpose_x = var_621_transpose_x_0, transpose_y = var_621_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_621")]; + tensor var_622 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_622")]; + tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 1, 4, 1])]; + tensor var_624 = reshape(shape = var_623, x = var_622)[name = tensor("op_624")]; + tensor cross_5 = mul(x = var_621, y = var_624)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_627 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_627")]; + tensor var_629_transpose_x_1 = const()[name = tensor("op_629_transpose_x_1"), val = tensor(true)]; + tensor var_629_transpose_y_1 = const()[name = tensor("op_629_transpose_y_1"), val = tensor(false)]; + tensor var_629 = matmul(transpose_x = var_629_transpose_x_1, transpose_y = var_629_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_629")]; + tensor new_kv_unnorm_5 = add(x = var_627, y = var_629)[name = tensor("new_kv_unnorm_5")]; + tensor var_631 = const()[name = tensor("op_631"), val = tensor(0x1p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_631)[name = tensor("new_scale_5")]; + tensor var_633 = sqrt(x = new_scale_5)[name = tensor("op_633")]; + tensor var_634 = real_div(x = new_kv_unnorm_5, y = var_633)[name = tensor("op_634")]; + tensor var_635_perm_0 = const()[name = tensor("op_635_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_635 = transpose(perm = var_635_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_635)[name = tensor("out_15")]; + tensor var_639 = const()[name = tensor("op_639"), val = tensor([1, 4, 256])]; + tensor out_17 = reshape(shape = var_639, x = out_15)[name = tensor("out_17")]; + tensor var_641 = silu(x = input_97)[name = tensor("op_641")]; + tensor input_99 = mul(x = var_641, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_21_begin_0 = const()[name = tensor("window_21_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_21_end_0 = const()[name = tensor("window_21_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_21_end_mask_0 = const()[name = tensor("window_21_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_21_squeeze_mask_0 = const()[name = tensor("window_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_21 = slice_by_index(begin = window_21_begin_0, end = window_21_end_0, end_mask = window_21_end_mask_0, squeeze_mask = window_21_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_21")]; + tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 0, 0])]; + tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; + tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, false, true])]; + tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; + tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; + tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; + tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; + tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; + tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; + tensor var_657_begin_0 = const()[name = tensor("op_657_begin_0"), val = tensor([0, 1, 0])]; + tensor var_657_end_0 = const()[name = tensor("op_657_end_0"), val = tensor([1, 2, 256])]; + tensor var_657_end_mask_0 = const()[name = tensor("op_657_end_mask_0"), val = tensor([true, false, true])]; + tensor var_657 = slice_by_index(begin = var_657_begin_0, end = var_657_end_0, end_mask = var_657_end_mask_0, x = x_15)[name = tensor("op_657")]; + tensor var_660_begin_0 = const()[name = tensor("op_660_begin_0"), val = tensor([0, 1, 0])]; + tensor var_660_end_0 = const()[name = tensor("op_660_end_0"), val = tensor([1, 16, 256])]; + tensor var_660_end_mask_0 = const()[name = tensor("op_660_end_mask_0"), val = tensor([true, true, true])]; + tensor var_660 = slice_by_index(begin = var_660_begin_0, end = var_660_end_0, end_mask = var_660_end_mask_0, x = window_23)[name = tensor("op_660")]; + tensor window_25_interleave_0 = const()[name = tensor("window_25_interleave_0"), val = tensor(false)]; + tensor window_25 = concat(axis = var_26, interleave = window_25_interleave_0, values = (var_660, var_657))[name = tensor("window_25")]; + tensor var_665_begin_0 = const()[name = tensor("op_665_begin_0"), val = tensor([0, 2, 0])]; + tensor var_665_end_0 = const()[name = tensor("op_665_end_0"), val = tensor([1, 3, 256])]; + tensor var_665_end_mask_0 = const()[name = tensor("op_665_end_mask_0"), val = tensor([true, false, true])]; + tensor var_665 = slice_by_index(begin = var_665_begin_0, end = var_665_end_0, end_mask = var_665_end_mask_0, x = x_15)[name = tensor("op_665")]; + tensor var_668_begin_0 = const()[name = tensor("op_668_begin_0"), val = tensor([0, 1, 0])]; + tensor var_668_end_0 = const()[name = tensor("op_668_end_0"), val = tensor([1, 16, 256])]; + tensor var_668_end_mask_0 = const()[name = tensor("op_668_end_mask_0"), val = tensor([true, true, true])]; + tensor var_668 = slice_by_index(begin = var_668_begin_0, end = var_668_end_0, end_mask = var_668_end_mask_0, x = window_25)[name = tensor("op_668")]; + tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; + tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_668, var_665))[name = tensor("window_27")]; + tensor var_673_begin_0 = const()[name = tensor("op_673_begin_0"), val = tensor([0, 3, 0])]; + tensor var_673_end_0 = const()[name = tensor("op_673_end_0"), val = tensor([1, 1, 256])]; + tensor var_673_end_mask_0 = const()[name = tensor("op_673_end_mask_0"), val = tensor([true, true, true])]; + tensor var_673 = slice_by_index(begin = var_673_begin_0, end = var_673_end_0, end_mask = var_673_end_mask_0, x = x_15)[name = tensor("op_673")]; + tensor var_676_begin_0 = const()[name = tensor("op_676_begin_0"), val = tensor([0, 1, 0])]; + tensor var_676_end_0 = const()[name = tensor("op_676_end_0"), val = tensor([1, 16, 256])]; + tensor var_676_end_mask_0 = const()[name = tensor("op_676_end_mask_0"), val = tensor([true, true, true])]; + tensor var_676 = slice_by_index(begin = var_676_begin_0, end = var_676_end_0, end_mask = var_676_end_mask_0, x = window_27)[name = tensor("op_676")]; + tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; + tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_676, var_673))[name = tensor("window_29")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_23, window_25, window_27, window_29))[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_701_split_sizes_0 = const()[name = tensor("op_701_split_sizes_0"), val = tensor([256, 256])]; + tensor var_701_axis_0 = const()[name = tensor("op_701_axis_0"), val = tensor(1)]; + tensor var_701_0, tensor var_701_1 = split(axis = var_701_axis_0, split_sizes = var_701_split_sizes_0, x = inputs_23)[name = tensor("op_701")]; + tensor var_703 = sigmoid(x = var_701_1)[name = tensor("op_703")]; + tensor inputs_25 = mul(x = var_701_0, y = var_703)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([4, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_734_begin_0 = const()[name = tensor("op_734_begin_0"), val = tensor([0, -1, 0])]; + tensor var_734_end_0 = const()[name = tensor("op_734_end_0"), val = tensor([4, 16, 256])]; + tensor var_734_end_mask_0 = const()[name = tensor("op_734_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_734 = slice_by_index(begin = var_734_begin_0, end = var_734_end_0, end_mask = var_734_end_mask_0, x = conv_out_5)[name = tensor("op_734")]; + tensor var_736_perm_0 = const()[name = tensor("op_736_perm_0"), val = tensor([1, 0, 2])]; + tensor var_736 = transpose(perm = var_736_perm_0, x = var_734)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_736)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_759 = const()[name = tensor("op_759"), val = tensor(0x1p-1)]; + tensor var_760 = mul(x = input_119, y = var_759)[name = tensor("op_760")]; + tensor input_121 = add(x = var_760, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_789 = const()[name = tensor("op_789"), val = tensor(0x1p-1)]; + tensor var_790 = mul(x = input_131, y = var_789)[name = tensor("op_790")]; + tensor input_133 = add(x = var_790, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_804 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 4, 4, 64])]; + tensor var_806 = reshape(shape = var_805, x = var_804)[name = tensor("op_806")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_810 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_811 = const()[name = tensor("op_811"), val = tensor(0x1p-3)]; + tensor var_812 = mul(x = var_810, y = var_811)[name = tensor("op_812")]; + tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 4, 4, 64])]; + tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_818 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_819 = const()[name = tensor("op_819"), val = tensor([1, 4, 4, 64])]; + tensor var_820 = reshape(shape = var_819, x = var_818)[name = tensor("op_820")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_814)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_806)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_830 = const()[name = tensor("op_830"), val = tensor([4, 1])]; + tensor var_831 = reshape(shape = var_830, x = sqrt_s_t_7)[name = tensor("op_831")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_831)[name = tensor("M_7")]; + tensor var_833 = mul(x = qk_7, y = M_7)[name = tensor("op_833")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_820)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_833, y = v_7)[name = tensor("inner_7")]; + tensor var_835_transpose_x_0 = const()[name = tensor("op_835_transpose_x_0"), val = tensor(false)]; + tensor var_835_transpose_y_0 = const()[name = tensor("op_835_transpose_y_0"), val = tensor(false)]; + tensor var_835 = matmul(transpose_x = var_835_transpose_x_0, transpose_y = var_835_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_835")]; + tensor var_836 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_836")]; + tensor var_837 = const()[name = tensor("op_837"), val = tensor([1, 1, 4, 1])]; + tensor var_838 = reshape(shape = var_837, x = var_836)[name = tensor("op_838")]; + tensor cross_7 = mul(x = var_835, y = var_838)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_841 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_841")]; + tensor var_843_transpose_x_1 = const()[name = tensor("op_843_transpose_x_1"), val = tensor(true)]; + tensor var_843_transpose_y_1 = const()[name = tensor("op_843_transpose_y_1"), val = tensor(false)]; + tensor var_843 = matmul(transpose_x = var_843_transpose_x_1, transpose_y = var_843_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_843")]; + tensor new_kv_unnorm_7 = add(x = var_841, y = var_843)[name = tensor("new_kv_unnorm_7")]; + tensor var_845 = const()[name = tensor("op_845"), val = tensor(0x1p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_845)[name = tensor("new_scale_7")]; + tensor var_847 = sqrt(x = new_scale_7)[name = tensor("op_847")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_847)[name = tensor("nkv_1")]; + tensor var_849_perm_0 = const()[name = tensor("op_849_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_849 = transpose(perm = var_849_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_849)[name = tensor("out_21")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 4, 256])]; + tensor out_23 = reshape(shape = var_853, x = out_21)[name = tensor("out_23")]; + tensor var_855 = silu(x = input_137)[name = tensor("op_855")]; + tensor input_139 = mul(x = var_855, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_31_begin_0 = const()[name = tensor("window_31_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_31_end_0 = const()[name = tensor("window_31_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_31_end_mask_0 = const()[name = tensor("window_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_31_squeeze_mask_0 = const()[name = tensor("window_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_31 = slice_by_index(begin = window_31_begin_0, end = window_31_end_0, end_mask = window_31_end_mask_0, squeeze_mask = window_31_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_31")]; + tensor var_863_begin_0 = const()[name = tensor("op_863_begin_0"), val = tensor([0, 0, 0])]; + tensor var_863_end_0 = const()[name = tensor("op_863_end_0"), val = tensor([1, 1, 256])]; + tensor var_863_end_mask_0 = const()[name = tensor("op_863_end_mask_0"), val = tensor([true, false, true])]; + tensor var_863 = slice_by_index(begin = var_863_begin_0, end = var_863_end_0, end_mask = var_863_end_mask_0, x = x_21)[name = tensor("op_863")]; + tensor var_866_begin_0 = const()[name = tensor("op_866_begin_0"), val = tensor([0, 1, 0])]; + tensor var_866_end_0 = const()[name = tensor("op_866_end_0"), val = tensor([1, 16, 256])]; + tensor var_866_end_mask_0 = const()[name = tensor("op_866_end_mask_0"), val = tensor([true, true, true])]; + tensor var_866 = slice_by_index(begin = var_866_begin_0, end = var_866_end_0, end_mask = var_866_end_mask_0, x = window_31)[name = tensor("op_866")]; + tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; + tensor window_33 = concat(axis = var_26, interleave = window_33_interleave_0, values = (var_866, var_863))[name = tensor("window_33")]; + tensor var_871_begin_0 = const()[name = tensor("op_871_begin_0"), val = tensor([0, 1, 0])]; + tensor var_871_end_0 = const()[name = tensor("op_871_end_0"), val = tensor([1, 2, 256])]; + tensor var_871_end_mask_0 = const()[name = tensor("op_871_end_mask_0"), val = tensor([true, false, true])]; + tensor var_871 = slice_by_index(begin = var_871_begin_0, end = var_871_end_0, end_mask = var_871_end_mask_0, x = x_21)[name = tensor("op_871")]; + tensor var_874_begin_0 = const()[name = tensor("op_874_begin_0"), val = tensor([0, 1, 0])]; + tensor var_874_end_0 = const()[name = tensor("op_874_end_0"), val = tensor([1, 16, 256])]; + tensor var_874_end_mask_0 = const()[name = tensor("op_874_end_mask_0"), val = tensor([true, true, true])]; + tensor var_874 = slice_by_index(begin = var_874_begin_0, end = var_874_end_0, end_mask = var_874_end_mask_0, x = window_33)[name = tensor("op_874")]; + tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; + tensor window_35 = concat(axis = var_26, interleave = window_35_interleave_0, values = (var_874, var_871))[name = tensor("window_35")]; + tensor var_879_begin_0 = const()[name = tensor("op_879_begin_0"), val = tensor([0, 2, 0])]; + tensor var_879_end_0 = const()[name = tensor("op_879_end_0"), val = tensor([1, 3, 256])]; + tensor var_879_end_mask_0 = const()[name = tensor("op_879_end_mask_0"), val = tensor([true, false, true])]; + tensor var_879 = slice_by_index(begin = var_879_begin_0, end = var_879_end_0, end_mask = var_879_end_mask_0, x = x_21)[name = tensor("op_879")]; + tensor var_882_begin_0 = const()[name = tensor("op_882_begin_0"), val = tensor([0, 1, 0])]; + tensor var_882_end_0 = const()[name = tensor("op_882_end_0"), val = tensor([1, 16, 256])]; + tensor var_882_end_mask_0 = const()[name = tensor("op_882_end_mask_0"), val = tensor([true, true, true])]; + tensor var_882 = slice_by_index(begin = var_882_begin_0, end = var_882_end_0, end_mask = var_882_end_mask_0, x = window_35)[name = tensor("op_882")]; + tensor window_37_interleave_0 = const()[name = tensor("window_37_interleave_0"), val = tensor(false)]; + tensor window_37 = concat(axis = var_26, interleave = window_37_interleave_0, values = (var_882, var_879))[name = tensor("window_37")]; + tensor var_887_begin_0 = const()[name = tensor("op_887_begin_0"), val = tensor([0, 3, 0])]; + tensor var_887_end_0 = const()[name = tensor("op_887_end_0"), val = tensor([1, 1, 256])]; + tensor var_887_end_mask_0 = const()[name = tensor("op_887_end_mask_0"), val = tensor([true, true, true])]; + tensor var_887 = slice_by_index(begin = var_887_begin_0, end = var_887_end_0, end_mask = var_887_end_mask_0, x = x_21)[name = tensor("op_887")]; + tensor var_890_begin_0 = const()[name = tensor("op_890_begin_0"), val = tensor([0, 1, 0])]; + tensor var_890_end_0 = const()[name = tensor("op_890_end_0"), val = tensor([1, 16, 256])]; + tensor var_890_end_mask_0 = const()[name = tensor("op_890_end_mask_0"), val = tensor([true, true, true])]; + tensor var_890 = slice_by_index(begin = var_890_begin_0, end = var_890_end_0, end_mask = var_890_end_mask_0, x = window_37)[name = tensor("op_890")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_890, var_887))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_33, window_35, window_37, window))[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_915_split_sizes_0 = const()[name = tensor("op_915_split_sizes_0"), val = tensor([256, 256])]; + tensor var_915_axis_0 = const()[name = tensor("op_915_axis_0"), val = tensor(1)]; + tensor var_915_0, tensor var_915_1 = split(axis = var_915_axis_0, split_sizes = var_915_split_sizes_0, x = inputs_33)[name = tensor("op_915")]; + tensor var_917 = sigmoid(x = var_915_1)[name = tensor("op_917")]; + tensor inputs_35 = mul(x = var_915_0, y = var_917)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([4, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_948_begin_0 = const()[name = tensor("op_948_begin_0"), val = tensor([0, -1, 0])]; + tensor var_948_end_0 = const()[name = tensor("op_948_end_0"), val = tensor([4, 16, 256])]; + tensor var_948_end_mask_0 = const()[name = tensor("op_948_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_948 = slice_by_index(begin = var_948_begin_0, end = var_948_end_0, end_mask = var_948_end_mask_0, x = conv_out_7)[name = tensor("op_948")]; + tensor var_950_perm_0 = const()[name = tensor("op_950_perm_0"), val = tensor([1, 0, 2])]; + tensor var_950 = transpose(perm = var_950_perm_0, x = var_948)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_950)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_973 = const()[name = tensor("op_973"), val = tensor(0x1p-1)]; + tensor var_974 = mul(x = input_159, y = var_973)[name = tensor("op_974")]; + tensor input_161 = add(x = var_974, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_992_begin_0 = const()[name = tensor("op_992_begin_0"), val = tensor([0, 0, 4])]; + tensor var_992_end_0 = const()[name = tensor("op_992_end_0"), val = tensor([1, 256, 22])]; + tensor var_992_end_mask_0 = const()[name = tensor("op_992_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_992_begin_0, end = var_992_end_0, end_mask = var_992_end_mask_0, x = cat)[name = tensor("op_992")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_994 = const()[name = tensor("op_994"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_995 = reduce_l2_norm(axes = var_994, keep_dims = var_29, x = input_163)[name = tensor("op_995")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_995)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_999_axis_0 = const()[name = tensor("op_999_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_999_axis_0, values = (var_206, var_420, var_634, nkv_1))[name = tensor("op_999")]; + tensor var_1001_axis_0 = const()[name = tensor("op_1001_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1001_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1001")]; + tensor var_1003_axis_0 = const()[name = tensor("op_1003_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1003_axis_0, values = (window_9, window_19, window_29, window))[name = tensor("op_1003")]; + tensor var_1012 = const()[name = tensor("op_1012"), val = tensor(0x1.5798eep-27)]; + tensor var_1017 = const()[name = tensor("op_1017"), val = tensor(0x1.4f8b58p-17)]; + tensor var_1019 = const()[name = tensor("op_1019"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_1020 = const()[name = tensor("op_1020"), val = tensor(true)]; + tensor var_1022 = const()[name = tensor("op_1022"), val = tensor(0x1p+0)]; + tensor var_1026 = const()[name = tensor("op_1026"), val = tensor(-1)]; + tensor var_1032 = const()[name = tensor("op_1032"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395584)))]; + tensor var_1094_axes_0 = const()[name = tensor("op_1094_axes_0"), val = tensor([2])]; + tensor var_1094 = expand_dims(axes = var_1094_axes_0, x = emb)[name = tensor("op_1094")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 9, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1094)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_1026, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1102_perm_0 = const()[name = tensor("op_1102_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([9, 4, 256])]; + tensor var_1102 = transpose(perm = var_1102_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1106, x = var_1102)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 9, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1114 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1115 = const()[name = tensor("op_1115"), val = tensor([9, 4, 4, 64])]; + tensor var_1116 = reshape(shape = var_1115, x = var_1114)[name = tensor("op_1116")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1120 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1121 = const()[name = tensor("op_1121"), val = tensor(0x1p-3)]; + tensor var_1122 = mul(x = var_1120, y = var_1121)[name = tensor("op_1122")]; + tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([9, 4, 4, 64])]; + tensor var_1124 = reshape(shape = var_1123, x = var_1122)[name = tensor("op_1124")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1128 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1129 = const()[name = tensor("op_1129"), val = tensor([9, 4, 4, 64])]; + tensor var_1130 = reshape(shape = var_1129, x = var_1128)[name = tensor("op_1130")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_1032, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_1022, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1124)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1116)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([1, 4])]; + tensor var_1143 = reshape(shape = var_1142, x = valid_mask)[name = tensor("op_1143")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1143)[name = tensor("causal_with_valid_1")]; + tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([4, 1])]; + tensor var_1146 = reshape(shape = var_1145, x = sqrt_s_t_9)[name = tensor("op_1146")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1146)[name = tensor("M_9")]; + tensor var_1148 = mul(x = qk_9, y = M_9)[name = tensor("op_1148")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1130)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1148, y = v_9)[name = tensor("inner_9")]; + tensor var_1150_transpose_x_0 = const()[name = tensor("op_1150_transpose_x_0"), val = tensor(false)]; + tensor var_1150_transpose_y_0 = const()[name = tensor("op_1150_transpose_y_0"), val = tensor(false)]; + tensor var_1150 = matmul(transpose_x = var_1150_transpose_x_0, transpose_y = var_1150_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1150")]; + tensor var_1151 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1151")]; + tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 1, 4, 1])]; + tensor var_1153 = reshape(shape = var_1152, x = var_1151)[name = tensor("op_1153")]; + tensor cross_9 = mul(x = var_1150, y = var_1153)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 1, 4, 1])]; + tensor var_1157 = reshape(shape = var_1156, x = valid_mask)[name = tensor("op_1157")]; + tensor v_masked_1 = mul(x = v_9, y = var_1157)[name = tensor("v_masked_1")]; + tensor var_1159 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1159")]; + tensor var_1161_transpose_x_1 = const()[name = tensor("op_1161_transpose_x_1"), val = tensor(true)]; + tensor var_1161_transpose_y_1 = const()[name = tensor("op_1161_transpose_y_1"), val = tensor(false)]; + tensor var_1161 = matmul(transpose_x = var_1161_transpose_x_1, transpose_y = var_1161_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1161")]; + tensor new_kv_unnorm_9 = add(x = var_1159, y = var_1161)[name = tensor("new_kv_unnorm_9")]; + tensor var_1163_keep_dims_0 = const()[name = tensor("op_1163_keep_dims_0"), val = tensor(false)]; + tensor var_1163 = reduce_sum(keep_dims = var_1163_keep_dims_0, x = valid_mask)[name = tensor("op_1163")]; + tensor var_1164 = const()[name = tensor("op_1164"), val = tensor([1])]; + tensor var_1165 = reshape(shape = var_1164, x = var_1163)[name = tensor("op_1165")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1165)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_1022, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1169 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1169")]; + tensor var_1170_perm_0 = const()[name = tensor("op_1170_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1170 = transpose(perm = var_1170_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1019, x = var_1170)[name = tensor("out_27")]; + tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([9, 4, 256])]; + tensor out_29 = reshape(shape = var_1174, x = out_27)[name = tensor("out_29")]; + tensor var_1176 = silu(x = input_169)[name = tensor("op_1176")]; + tensor input_171 = mul(x = var_1176, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1017, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1186 = const()[name = tensor("op_1186"), val = tensor([1, 9, 4, 256])]; + tensor var_1187 = reshape(shape = var_1186, x = xt_1)[name = tensor("op_1187")]; + tensor var_1188_perm_0 = const()[name = tensor("op_1188_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([4, 9, 256])]; + tensor var_1188 = transpose(perm = var_1188_perm_0, x = var_1187)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1191, x = var_1188)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1214 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([9, 4, 3, 256])]; + tensor var_1216 = reshape(shape = concat_1, x = var_1214)[name = tensor("op_1216")]; + tensor var_1217_axes_0 = const()[name = tensor("op_1217_axes_0"), val = tensor([0])]; + tensor var_1217 = expand_dims(axes = var_1217_axes_0, x = var_1216)[name = tensor("op_1217")]; + tensor var_1218_perm_0 = const()[name = tensor("op_1218_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1219_axes_0 = const()[name = tensor("op_1219_axes_0"), val = tensor([-2])]; + tensor var_1218 = transpose(perm = var_1218_perm_0, x = var_1217)[name = tensor("transpose_21")]; + tensor var_1219 = squeeze(axes = var_1219_axes_0, x = var_1218)[name = tensor("op_1219")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 9, 4, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1219)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 9, 4, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1219)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 9, 4, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1219)[name = tensor("v_11")]; + tensor var_1227 = const()[name = tensor("op_1227"), val = tensor([9, 16, 64])]; + tensor var_1228 = reshape(shape = var_1227, x = q_11)[name = tensor("op_1228")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1234 = const()[name = tensor("op_1234"), val = tensor([9, 16, 64])]; + tensor var_1235 = reshape(shape = var_1234, x = k_11)[name = tensor("op_1235")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1241 = const()[name = tensor("op_1241"), val = tensor([9, 16, 64])]; + tensor var_1242 = reshape(shape = var_1241, x = v_11)[name = tensor("op_1242")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([4, 4, 9, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1228)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1245, x = q_13)[name = tensor("q_15")]; + tensor var_1247 = const()[name = tensor("op_1247"), val = tensor([4, 4, 9, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1235)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1247, x = k_13)[name = tensor("k_15")]; + tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([4, 4, 9, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1242)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1249, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([2, 0, 1, 3])]; + tensor var_1257 = const()[name = tensor("op_1257"), val = tensor([36, 256])]; + tensor var_1253 = transpose(perm = var_1252, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1257, x = var_1253)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([9, 4, 256])]; + tensor attn_output_7 = reshape(shape = var_1261, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1017, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1017, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1281 = const()[name = tensor("op_1281"), val = tensor([1, 4, 9, 256])]; + tensor x_31 = reshape(shape = var_1281, x = xt_3)[name = tensor("x_31")]; + tensor var_1283_perm_0 = const()[name = tensor("op_1283_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1287 = const()[name = tensor("op_1287"), val = tensor([9, 4, 256])]; + tensor var_1283 = transpose(perm = var_1283_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1287, x = var_1283)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 9, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1295 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1296 = const()[name = tensor("op_1296"), val = tensor([9, 4, 4, 64])]; + tensor var_1297 = reshape(shape = var_1296, x = var_1295)[name = tensor("op_1297")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1301 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1302 = const()[name = tensor("op_1302"), val = tensor(0x1p-3)]; + tensor var_1303 = mul(x = var_1301, y = var_1302)[name = tensor("op_1303")]; + tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([9, 4, 4, 64])]; + tensor var_1305 = reshape(shape = var_1304, x = var_1303)[name = tensor("op_1305")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1309 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1310 = const()[name = tensor("op_1310"), val = tensor([9, 4, 4, 64])]; + tensor var_1311 = reshape(shape = var_1310, x = var_1309)[name = tensor("op_1311")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_1022, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1305)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1297)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([4, 1])]; + tensor var_1327 = reshape(shape = var_1326, x = sqrt_s_t)[name = tensor("op_1327")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1327)[name = tensor("M")]; + tensor var_1329 = mul(x = qk, y = M)[name = tensor("op_1329")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1311)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1329, y = v_17)[name = tensor("inner")]; + tensor var_1331_transpose_x_0 = const()[name = tensor("op_1331_transpose_x_0"), val = tensor(false)]; + tensor var_1331_transpose_y_0 = const()[name = tensor("op_1331_transpose_y_0"), val = tensor(false)]; + tensor var_1331 = matmul(transpose_x = var_1331_transpose_x_0, transpose_y = var_1331_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1331")]; + tensor var_1332 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1332")]; + tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 1, 4, 1])]; + tensor var_1334 = reshape(shape = var_1333, x = var_1332)[name = tensor("op_1334")]; + tensor cross = mul(x = var_1331, y = var_1334)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1157)[name = tensor("v_masked")]; + tensor var_1340 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1340")]; + tensor var_1342_transpose_x_1 = const()[name = tensor("op_1342_transpose_x_1"), val = tensor(true)]; + tensor var_1342_transpose_y_1 = const()[name = tensor("op_1342_transpose_y_1"), val = tensor(false)]; + tensor var_1342 = matmul(transpose_x = var_1342_transpose_x_1, transpose_y = var_1342_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1342")]; + tensor new_kv_unnorm = add(x = var_1340, y = var_1342)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1165)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_1022, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1351_perm_0 = const()[name = tensor("op_1351_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1351 = transpose(perm = var_1351_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1019, x = var_1351)[name = tensor("out_33")]; + tensor var_1355 = const()[name = tensor("op_1355"), val = tensor([9, 4, 256])]; + tensor out = reshape(shape = var_1355, x = out_33)[name = tensor("out")]; + tensor var_1357 = silu(x = input_187)[name = tensor("op_1357")]; + tensor input_189 = mul(x = var_1357, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1017, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1367 = const()[name = tensor("op_1367"), val = tensor([1, 9, 4, 256])]; + tensor var_1368 = reshape(shape = var_1367, x = xt_5)[name = tensor("op_1368")]; + tensor var_1369_perm_0 = const()[name = tensor("op_1369_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1372 = const()[name = tensor("op_1372"), val = tensor([4, 9, 256])]; + tensor var_1369 = transpose(perm = var_1369_perm_0, x = var_1368)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1372, x = var_1369)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1395 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([9, 4, 3, 256])]; + tensor var_1397 = reshape(shape = concat_2, x = var_1395)[name = tensor("op_1397")]; + tensor var_1398_axes_0 = const()[name = tensor("op_1398_axes_0"), val = tensor([0])]; + tensor var_1398 = expand_dims(axes = var_1398_axes_0, x = var_1397)[name = tensor("op_1398")]; + tensor var_1399_perm_0 = const()[name = tensor("op_1399_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1400_axes_0 = const()[name = tensor("op_1400_axes_0"), val = tensor([-2])]; + tensor var_1399 = transpose(perm = var_1399_perm_0, x = var_1398)[name = tensor("transpose_8")]; + tensor var_1400 = squeeze(axes = var_1400_axes_0, x = var_1399)[name = tensor("op_1400")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 9, 4, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1400)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 9, 4, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1400)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 9, 4, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1400)[name = tensor("v_19")]; + tensor var_1408 = const()[name = tensor("op_1408"), val = tensor([9, 16, 64])]; + tensor var_1409 = reshape(shape = var_1408, x = q_19)[name = tensor("op_1409")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1415 = const()[name = tensor("op_1415"), val = tensor([9, 16, 64])]; + tensor var_1416 = reshape(shape = var_1415, x = k_19)[name = tensor("op_1416")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1422 = const()[name = tensor("op_1422"), val = tensor([9, 16, 64])]; + tensor var_1423 = reshape(shape = var_1422, x = v_19)[name = tensor("op_1423")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1426 = const()[name = tensor("op_1426"), val = tensor([4, 4, 9, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1409)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1426, x = q_21)[name = tensor("q")]; + tensor var_1428 = const()[name = tensor("op_1428"), val = tensor([4, 4, 9, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1416)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1428, x = k_21)[name = tensor("k")]; + tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([4, 4, 9, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1423)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1430, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([2, 0, 1, 3])]; + tensor var_1438 = const()[name = tensor("op_1438"), val = tensor([36, 256])]; + tensor var_1434 = transpose(perm = var_1433, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1438, x = var_1434)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([9, 4, 256])]; + tensor attn_output = reshape(shape = var_1442, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1017, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1017, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1462 = const()[name = tensor("op_1462"), val = tensor([1, 4, 9, 256])]; + tensor input = reshape(shape = var_1462, x = xt)[name = tensor("input")]; + tensor var_1464 = const()[name = tensor("op_1464"), val = tensor([-1])]; + tensor var_1465 = reduce_l2_norm(axes = var_1464, keep_dims = var_1020, x = input)[name = tensor("op_1465")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_1012, beta = const_42, x = var_1465)[name = tensor("clip_5")]; + tensor var_1467 = real_div(x = input, y = clip_5)[name = tensor("op_1467")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([4, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([4, 256, 9])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1467)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 4, 9])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 4, 8])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1471")]; + tensor var_1473_axis_0 = const()[name = tensor("op_1473_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1473_axis_0, values = (var_1169, nkv))[name = tensor("op_1473")]; + tensor var_1475_axis_0 = const()[name = tensor("op_1475_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1475_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1475")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/weights/weight.bin b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..bb752b5ff6aa2f51faf29649b929db88f6f15bed --- /dev/null +++ b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e9b5f77e4a310fc720a41d6b080ba7693bba06a3accd2c66bea0fc835f00d9b +size 44432512 diff --git a/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..adac95464c7b8d573c30be0062638e5c2c0320d8 --- /dev/null +++ b/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:182619e9c96f802ccae152cbda7d15ae2b22f480c04c689709f70aafdc13e145 +size 191044 diff --git a/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..bb752b5ff6aa2f51faf29649b929db88f6f15bed --- /dev/null +++ b/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e9b5f77e4a310fc720a41d6b080ba7693bba06a3accd2c66bea0fc835f00d9b +size 44432512 diff --git a/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Manifest.json b/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..36202097fcb62b8f48444015affe7358861ad8e4 --- /dev/null +++ b/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "1C991CEA-BA44-4416-8759-2F7BBCD14BF6": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "676FBDCA-C6C7-4101-9A73-E68F17B5F2F1": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "676FBDCA-C6C7-4101-9A73-E68F17B5F2F1" +} diff --git a/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/analytics/coremldata.bin b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..3126a415e64304c0c88002baacbed295e38ad760 --- /dev/null +++ b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:631a472bbce1c64c900ce12bd152385af02044081a60d2b7634e29c403b9bc69 +size 243 diff --git a/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/coremldata.bin b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..54dd550a42c104727ee2b06639b7b4ba3bc7bf83 --- /dev/null +++ b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3cce9489e7b9d36fdbfd8fa6395dc2943f598f22d4ce2b28f349fff244c15703 +size 1301 diff --git a/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/metadata.json b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..6c6b1d83c6b133bc9631ee243751876ec6e39cef --- /dev/null +++ b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND CALLHOME streaming diarizer (pipeline, T=5, max_speakers=7)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 5 × 7)", + "shortDescription" : "", + "shape" : "[1, 5, 7]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 9 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 9, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 68, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 72, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 26, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 47 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 5 × 345)", + "shortDescription" : "", + "shape" : "[1, 5, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 9 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 9, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 5)", + "shortDescription" : "", + "shape" : "[5]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"ch\", \"model_label\": \"CALLHOME\", \"variant\": \"pipeline\", \"chunk_size\": 5, \"step_duration_ms\": 500, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 7, \"max_nspks\": 9, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_ch_500ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/model.mil b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..fce20b41f60d6b106fa6a87a31e58eb7e064dc8b --- /dev/null +++ b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/model.mil @@ -0,0 +1,1366 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2, 0x1.4p+2])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982144)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983232)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336576)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337664)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338752)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339840)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340928)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345088)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393728)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394816)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443456)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444544)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445632)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446720)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708928)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7710016)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972224)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973312)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235520)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236608)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498816)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499904)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762112)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763200)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764288)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766400)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290752)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307200)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308288)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309376)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310464)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311552)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312640)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574848)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575936)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9577024)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581184)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629824)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630912)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679552)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680640)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681728)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682816)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683904)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688064)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736704)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737792)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786432)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787520)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788608)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789696)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051904)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052992)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315200)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316288)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578496)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579584)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841792)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842880)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105088)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106176)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107264)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109376)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633728)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650176)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651264)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652352)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653440)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654528)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655616)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917824)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918912)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15920000)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924160)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972800)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973888)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022528)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023616)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024704)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025792)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026880)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18031040)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079680)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080768)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129408)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130496)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131584)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132672)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394880)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395968)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658176)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659264)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921472)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922560)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184768)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185856)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448064)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449152)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450240)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452352)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976704)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993152)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994240)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995328)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996416)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997504)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998592)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260800)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261888)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262976)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267136)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315776)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316864)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365504)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366592)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367680)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368768)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369856)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24374016)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422656)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423744)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472384)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473472)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474560)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475648)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737856)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738944)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001152)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002240)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264448)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265536)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527744)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528832)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791040)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792128)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793216)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795328)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319680)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336128)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337216)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338304)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339392)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340480)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341568)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603776)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604864)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605952)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610112)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658752)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659840)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708480)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709568)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710656)))]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710848)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711936)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236288)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237376)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499584)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500672)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762880)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763968)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026176)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027264)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289472)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290560)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552768)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553856)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554944)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32556032)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818240)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821376)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607872)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608960)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33610048)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618304)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715520)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716608)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813824)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814912)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816000)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37817088)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079296)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080384)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342592)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343680)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605888)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606976)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869184)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870272)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132480)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133568)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134656)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135744)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397952)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39401088)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187584)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188672)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189760)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40198016)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295232)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296320)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393536)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394624)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_18 = const()[name = tensor("op_18"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_21 = const()[name = tensor("op_21"), val = tensor(2)]; + tensor var_24 = const()[name = tensor("op_24"), val = tensor(0)]; + tensor var_27 = const()[name = tensor("op_27"), val = tensor(1)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(0x1.4f8b58p-17)]; + tensor var_30 = const()[name = tensor("op_30"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_29, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_148 = const()[name = tensor("op_148"), val = tensor(0x1p-1)]; + tensor var_149 = mul(x = input_11, y = var_148)[name = tensor("op_149")]; + tensor input_13 = add(x = var_149, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_29, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_163 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_164 = const()[name = tensor("op_164"), val = tensor([1, 5, 4, 64])]; + tensor var_165 = reshape(shape = var_164, x = var_163)[name = tensor("op_165")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_169 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_170 = const()[name = tensor("op_170"), val = tensor(0x1p-3)]; + tensor var_171 = mul(x = var_169, y = var_170)[name = tensor("op_171")]; + tensor var_172 = const()[name = tensor("op_172"), val = tensor([1, 5, 4, 64])]; + tensor var_173 = reshape(shape = var_172, x = var_171)[name = tensor("op_173")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_177 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_178 = const()[name = tensor("op_178"), val = tensor([1, 5, 4, 64])]; + tensor var_179 = reshape(shape = var_178, x = var_177)[name = tensor("op_179")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_173)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_165)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_189 = const()[name = tensor("op_189"), val = tensor([5, 1])]; + tensor var_190 = reshape(shape = var_189, x = sqrt_s_t_1)[name = tensor("op_190")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_190)[name = tensor("M_1")]; + tensor var_192 = mul(x = qk_1, y = M_1)[name = tensor("op_192")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_179)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_192, y = v_1)[name = tensor("inner_1")]; + tensor var_194_transpose_x_0 = const()[name = tensor("op_194_transpose_x_0"), val = tensor(false)]; + tensor var_194_transpose_y_0 = const()[name = tensor("op_194_transpose_y_0"), val = tensor(false)]; + tensor var_194 = matmul(transpose_x = var_194_transpose_x_0, transpose_y = var_194_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_194")]; + tensor var_195 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_195")]; + tensor var_196 = const()[name = tensor("op_196"), val = tensor([1, 1, 5, 1])]; + tensor var_197 = reshape(shape = var_196, x = var_195)[name = tensor("op_197")]; + tensor cross_1 = mul(x = var_194, y = var_197)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_200 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_200")]; + tensor var_202_transpose_x_1 = const()[name = tensor("op_202_transpose_x_1"), val = tensor(true)]; + tensor var_202_transpose_y_1 = const()[name = tensor("op_202_transpose_y_1"), val = tensor(false)]; + tensor var_202 = matmul(transpose_x = var_202_transpose_x_1, transpose_y = var_202_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_202")]; + tensor new_kv_unnorm_1 = add(x = var_200, y = var_202)[name = tensor("new_kv_unnorm_1")]; + tensor var_204 = const()[name = tensor("op_204"), val = tensor(0x1.4p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_204)[name = tensor("new_scale_1")]; + tensor var_206 = sqrt(x = new_scale_1)[name = tensor("op_206")]; + tensor var_207 = real_div(x = new_kv_unnorm_1, y = var_206)[name = tensor("op_207")]; + tensor var_208_perm_0 = const()[name = tensor("op_208_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_208 = transpose(perm = var_208_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_18, x = var_208)[name = tensor("out_3")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor([1, 5, 256])]; + tensor out_5 = reshape(shape = var_212, x = out_3)[name = tensor("out_5")]; + tensor var_214 = silu(x = input_17)[name = tensor("op_214")]; + tensor input_19 = mul(x = var_214, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_222_begin_0 = const()[name = tensor("op_222_begin_0"), val = tensor([0, 0, 0])]; + tensor var_222_end_0 = const()[name = tensor("op_222_end_0"), val = tensor([1, 1, 256])]; + tensor var_222_end_mask_0 = const()[name = tensor("op_222_end_mask_0"), val = tensor([true, false, true])]; + tensor var_222 = slice_by_index(begin = var_222_begin_0, end = var_222_end_0, end_mask = var_222_end_mask_0, x = x_3)[name = tensor("op_222")]; + tensor var_225_begin_0 = const()[name = tensor("op_225_begin_0"), val = tensor([0, 1, 0])]; + tensor var_225_end_0 = const()[name = tensor("op_225_end_0"), val = tensor([1, 16, 256])]; + tensor var_225_end_mask_0 = const()[name = tensor("op_225_end_mask_0"), val = tensor([true, true, true])]; + tensor var_225 = slice_by_index(begin = var_225_begin_0, end = var_225_end_0, end_mask = var_225_end_mask_0, x = window_1)[name = tensor("op_225")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_27, interleave = window_3_interleave_0, values = (var_225, var_222))[name = tensor("window_3")]; + tensor var_230_begin_0 = const()[name = tensor("op_230_begin_0"), val = tensor([0, 1, 0])]; + tensor var_230_end_0 = const()[name = tensor("op_230_end_0"), val = tensor([1, 2, 256])]; + tensor var_230_end_mask_0 = const()[name = tensor("op_230_end_mask_0"), val = tensor([true, false, true])]; + tensor var_230 = slice_by_index(begin = var_230_begin_0, end = var_230_end_0, end_mask = var_230_end_mask_0, x = x_3)[name = tensor("op_230")]; + tensor var_233_begin_0 = const()[name = tensor("op_233_begin_0"), val = tensor([0, 1, 0])]; + tensor var_233_end_0 = const()[name = tensor("op_233_end_0"), val = tensor([1, 16, 256])]; + tensor var_233_end_mask_0 = const()[name = tensor("op_233_end_mask_0"), val = tensor([true, true, true])]; + tensor var_233 = slice_by_index(begin = var_233_begin_0, end = var_233_end_0, end_mask = var_233_end_mask_0, x = window_3)[name = tensor("op_233")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_27, interleave = window_5_interleave_0, values = (var_233, var_230))[name = tensor("window_5")]; + tensor var_238_begin_0 = const()[name = tensor("op_238_begin_0"), val = tensor([0, 2, 0])]; + tensor var_238_end_0 = const()[name = tensor("op_238_end_0"), val = tensor([1, 3, 256])]; + tensor var_238_end_mask_0 = const()[name = tensor("op_238_end_mask_0"), val = tensor([true, false, true])]; + tensor var_238 = slice_by_index(begin = var_238_begin_0, end = var_238_end_0, end_mask = var_238_end_mask_0, x = x_3)[name = tensor("op_238")]; + tensor var_241_begin_0 = const()[name = tensor("op_241_begin_0"), val = tensor([0, 1, 0])]; + tensor var_241_end_0 = const()[name = tensor("op_241_end_0"), val = tensor([1, 16, 256])]; + tensor var_241_end_mask_0 = const()[name = tensor("op_241_end_mask_0"), val = tensor([true, true, true])]; + tensor var_241 = slice_by_index(begin = var_241_begin_0, end = var_241_end_0, end_mask = var_241_end_mask_0, x = window_5)[name = tensor("op_241")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_27, interleave = window_7_interleave_0, values = (var_241, var_238))[name = tensor("window_7")]; + tensor var_246_begin_0 = const()[name = tensor("op_246_begin_0"), val = tensor([0, 3, 0])]; + tensor var_246_end_0 = const()[name = tensor("op_246_end_0"), val = tensor([1, 4, 256])]; + tensor var_246_end_mask_0 = const()[name = tensor("op_246_end_mask_0"), val = tensor([true, false, true])]; + tensor var_246 = slice_by_index(begin = var_246_begin_0, end = var_246_end_0, end_mask = var_246_end_mask_0, x = x_3)[name = tensor("op_246")]; + tensor var_249_begin_0 = const()[name = tensor("op_249_begin_0"), val = tensor([0, 1, 0])]; + tensor var_249_end_0 = const()[name = tensor("op_249_end_0"), val = tensor([1, 16, 256])]; + tensor var_249_end_mask_0 = const()[name = tensor("op_249_end_mask_0"), val = tensor([true, true, true])]; + tensor var_249 = slice_by_index(begin = var_249_begin_0, end = var_249_end_0, end_mask = var_249_end_mask_0, x = window_7)[name = tensor("op_249")]; + tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; + tensor window_9 = concat(axis = var_27, interleave = window_9_interleave_0, values = (var_249, var_246))[name = tensor("window_9")]; + tensor var_254_begin_0 = const()[name = tensor("op_254_begin_0"), val = tensor([0, 4, 0])]; + tensor var_254_end_0 = const()[name = tensor("op_254_end_0"), val = tensor([1, 1, 256])]; + tensor var_254_end_mask_0 = const()[name = tensor("op_254_end_mask_0"), val = tensor([true, true, true])]; + tensor var_254 = slice_by_index(begin = var_254_begin_0, end = var_254_end_0, end_mask = var_254_end_mask_0, x = x_3)[name = tensor("op_254")]; + tensor var_257_begin_0 = const()[name = tensor("op_257_begin_0"), val = tensor([0, 1, 0])]; + tensor var_257_end_0 = const()[name = tensor("op_257_end_0"), val = tensor([1, 16, 256])]; + tensor var_257_end_mask_0 = const()[name = tensor("op_257_end_mask_0"), val = tensor([true, true, true])]; + tensor var_257 = slice_by_index(begin = var_257_begin_0, end = var_257_end_0, end_mask = var_257_end_mask_0, x = window_9)[name = tensor("op_257")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_27, interleave = window_11_interleave_0, values = (var_257, var_254))[name = tensor("window_11")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_24, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9, window_11))[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_282_split_sizes_0 = const()[name = tensor("op_282_split_sizes_0"), val = tensor([256, 256])]; + tensor var_282_axis_0 = const()[name = tensor("op_282_axis_0"), val = tensor(1)]; + tensor var_282_0, tensor var_282_1 = split(axis = var_282_axis_0, split_sizes = var_282_split_sizes_0, x = inputs_3)[name = tensor("op_282")]; + tensor var_284 = sigmoid(x = var_282_1)[name = tensor("op_284")]; + tensor inputs_5 = mul(x = var_282_0, y = var_284)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([5, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_315_begin_0 = const()[name = tensor("op_315_begin_0"), val = tensor([0, -1, 0])]; + tensor var_315_end_0 = const()[name = tensor("op_315_end_0"), val = tensor([5, 16, 256])]; + tensor var_315_end_mask_0 = const()[name = tensor("op_315_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_315 = slice_by_index(begin = var_315_begin_0, end = var_315_end_0, end_mask = var_315_end_mask_0, x = conv_out_1)[name = tensor("op_315")]; + tensor var_317_perm_0 = const()[name = tensor("op_317_perm_0"), val = tensor([1, 0, 2])]; + tensor var_317 = transpose(perm = var_317_perm_0, x = var_315)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_317)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor(0x1p-1)]; + tensor var_341 = mul(x = input_39, y = var_340)[name = tensor("op_341")]; + tensor input_41 = add(x = var_341, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_29, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_370 = const()[name = tensor("op_370"), val = tensor(0x1p-1)]; + tensor var_371 = mul(x = input_51, y = var_370)[name = tensor("op_371")]; + tensor input_53 = add(x = var_371, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_29, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_385 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([1, 5, 4, 64])]; + tensor var_387 = reshape(shape = var_386, x = var_385)[name = tensor("op_387")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_391 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_392 = const()[name = tensor("op_392"), val = tensor(0x1p-3)]; + tensor var_393 = mul(x = var_391, y = var_392)[name = tensor("op_393")]; + tensor var_394 = const()[name = tensor("op_394"), val = tensor([1, 5, 4, 64])]; + tensor var_395 = reshape(shape = var_394, x = var_393)[name = tensor("op_395")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_399 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_400 = const()[name = tensor("op_400"), val = tensor([1, 5, 4, 64])]; + tensor var_401 = reshape(shape = var_400, x = var_399)[name = tensor("op_401")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_395)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_387)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([5, 1])]; + tensor var_412 = reshape(shape = var_411, x = sqrt_s_t_3)[name = tensor("op_412")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_412)[name = tensor("M_3")]; + tensor var_414 = mul(x = qk_3, y = M_3)[name = tensor("op_414")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_401)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_414, y = v_3)[name = tensor("inner_3")]; + tensor var_416_transpose_x_0 = const()[name = tensor("op_416_transpose_x_0"), val = tensor(false)]; + tensor var_416_transpose_y_0 = const()[name = tensor("op_416_transpose_y_0"), val = tensor(false)]; + tensor var_416 = matmul(transpose_x = var_416_transpose_x_0, transpose_y = var_416_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_416")]; + tensor var_417 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_417")]; + tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1, 5, 1])]; + tensor var_419 = reshape(shape = var_418, x = var_417)[name = tensor("op_419")]; + tensor cross_3 = mul(x = var_416, y = var_419)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_422 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_422")]; + tensor var_424_transpose_x_1 = const()[name = tensor("op_424_transpose_x_1"), val = tensor(true)]; + tensor var_424_transpose_y_1 = const()[name = tensor("op_424_transpose_y_1"), val = tensor(false)]; + tensor var_424 = matmul(transpose_x = var_424_transpose_x_1, transpose_y = var_424_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_424")]; + tensor new_kv_unnorm_3 = add(x = var_422, y = var_424)[name = tensor("new_kv_unnorm_3")]; + tensor var_426 = const()[name = tensor("op_426"), val = tensor(0x1.4p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_426)[name = tensor("new_scale_3")]; + tensor var_428 = sqrt(x = new_scale_3)[name = tensor("op_428")]; + tensor var_429 = real_div(x = new_kv_unnorm_3, y = var_428)[name = tensor("op_429")]; + tensor var_430_perm_0 = const()[name = tensor("op_430_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_430 = transpose(perm = var_430_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_18, x = var_430)[name = tensor("out_9")]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 5, 256])]; + tensor out_11 = reshape(shape = var_434, x = out_9)[name = tensor("out_11")]; + tensor var_436 = silu(x = input_57)[name = tensor("op_436")]; + tensor input_59 = mul(x = var_436, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; + tensor var_444_begin_0 = const()[name = tensor("op_444_begin_0"), val = tensor([0, 0, 0])]; + tensor var_444_end_0 = const()[name = tensor("op_444_end_0"), val = tensor([1, 1, 256])]; + tensor var_444_end_mask_0 = const()[name = tensor("op_444_end_mask_0"), val = tensor([true, false, true])]; + tensor var_444 = slice_by_index(begin = var_444_begin_0, end = var_444_end_0, end_mask = var_444_end_mask_0, x = x_9)[name = tensor("op_444")]; + tensor var_447_begin_0 = const()[name = tensor("op_447_begin_0"), val = tensor([0, 1, 0])]; + tensor var_447_end_0 = const()[name = tensor("op_447_end_0"), val = tensor([1, 16, 256])]; + tensor var_447_end_mask_0 = const()[name = tensor("op_447_end_mask_0"), val = tensor([true, true, true])]; + tensor var_447 = slice_by_index(begin = var_447_begin_0, end = var_447_end_0, end_mask = var_447_end_mask_0, x = window_13)[name = tensor("op_447")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_27, interleave = window_15_interleave_0, values = (var_447, var_444))[name = tensor("window_15")]; + tensor var_452_begin_0 = const()[name = tensor("op_452_begin_0"), val = tensor([0, 1, 0])]; + tensor var_452_end_0 = const()[name = tensor("op_452_end_0"), val = tensor([1, 2, 256])]; + tensor var_452_end_mask_0 = const()[name = tensor("op_452_end_mask_0"), val = tensor([true, false, true])]; + tensor var_452 = slice_by_index(begin = var_452_begin_0, end = var_452_end_0, end_mask = var_452_end_mask_0, x = x_9)[name = tensor("op_452")]; + tensor var_455_begin_0 = const()[name = tensor("op_455_begin_0"), val = tensor([0, 1, 0])]; + tensor var_455_end_0 = const()[name = tensor("op_455_end_0"), val = tensor([1, 16, 256])]; + tensor var_455_end_mask_0 = const()[name = tensor("op_455_end_mask_0"), val = tensor([true, true, true])]; + tensor var_455 = slice_by_index(begin = var_455_begin_0, end = var_455_end_0, end_mask = var_455_end_mask_0, x = window_15)[name = tensor("op_455")]; + tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; + tensor window_17 = concat(axis = var_27, interleave = window_17_interleave_0, values = (var_455, var_452))[name = tensor("window_17")]; + tensor var_460_begin_0 = const()[name = tensor("op_460_begin_0"), val = tensor([0, 2, 0])]; + tensor var_460_end_0 = const()[name = tensor("op_460_end_0"), val = tensor([1, 3, 256])]; + tensor var_460_end_mask_0 = const()[name = tensor("op_460_end_mask_0"), val = tensor([true, false, true])]; + tensor var_460 = slice_by_index(begin = var_460_begin_0, end = var_460_end_0, end_mask = var_460_end_mask_0, x = x_9)[name = tensor("op_460")]; + tensor var_463_begin_0 = const()[name = tensor("op_463_begin_0"), val = tensor([0, 1, 0])]; + tensor var_463_end_0 = const()[name = tensor("op_463_end_0"), val = tensor([1, 16, 256])]; + tensor var_463_end_mask_0 = const()[name = tensor("op_463_end_mask_0"), val = tensor([true, true, true])]; + tensor var_463 = slice_by_index(begin = var_463_begin_0, end = var_463_end_0, end_mask = var_463_end_mask_0, x = window_17)[name = tensor("op_463")]; + tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; + tensor window_19 = concat(axis = var_27, interleave = window_19_interleave_0, values = (var_463, var_460))[name = tensor("window_19")]; + tensor var_468_begin_0 = const()[name = tensor("op_468_begin_0"), val = tensor([0, 3, 0])]; + tensor var_468_end_0 = const()[name = tensor("op_468_end_0"), val = tensor([1, 4, 256])]; + tensor var_468_end_mask_0 = const()[name = tensor("op_468_end_mask_0"), val = tensor([true, false, true])]; + tensor var_468 = slice_by_index(begin = var_468_begin_0, end = var_468_end_0, end_mask = var_468_end_mask_0, x = x_9)[name = tensor("op_468")]; + tensor var_471_begin_0 = const()[name = tensor("op_471_begin_0"), val = tensor([0, 1, 0])]; + tensor var_471_end_0 = const()[name = tensor("op_471_end_0"), val = tensor([1, 16, 256])]; + tensor var_471_end_mask_0 = const()[name = tensor("op_471_end_mask_0"), val = tensor([true, true, true])]; + tensor var_471 = slice_by_index(begin = var_471_begin_0, end = var_471_end_0, end_mask = var_471_end_mask_0, x = window_19)[name = tensor("op_471")]; + tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; + tensor window_21 = concat(axis = var_27, interleave = window_21_interleave_0, values = (var_471, var_468))[name = tensor("window_21")]; + tensor var_476_begin_0 = const()[name = tensor("op_476_begin_0"), val = tensor([0, 4, 0])]; + tensor var_476_end_0 = const()[name = tensor("op_476_end_0"), val = tensor([1, 1, 256])]; + tensor var_476_end_mask_0 = const()[name = tensor("op_476_end_mask_0"), val = tensor([true, true, true])]; + tensor var_476 = slice_by_index(begin = var_476_begin_0, end = var_476_end_0, end_mask = var_476_end_mask_0, x = x_9)[name = tensor("op_476")]; + tensor var_479_begin_0 = const()[name = tensor("op_479_begin_0"), val = tensor([0, 1, 0])]; + tensor var_479_end_0 = const()[name = tensor("op_479_end_0"), val = tensor([1, 16, 256])]; + tensor var_479_end_mask_0 = const()[name = tensor("op_479_end_mask_0"), val = tensor([true, true, true])]; + tensor var_479 = slice_by_index(begin = var_479_begin_0, end = var_479_end_0, end_mask = var_479_end_mask_0, x = window_21)[name = tensor("op_479")]; + tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; + tensor window_23 = concat(axis = var_27, interleave = window_23_interleave_0, values = (var_479, var_476))[name = tensor("window_23")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_24, interleave = input_61_interleave_0, values = (window_15, window_17, window_19, window_21, window_23))[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_504_split_sizes_0 = const()[name = tensor("op_504_split_sizes_0"), val = tensor([256, 256])]; + tensor var_504_axis_0 = const()[name = tensor("op_504_axis_0"), val = tensor(1)]; + tensor var_504_0, tensor var_504_1 = split(axis = var_504_axis_0, split_sizes = var_504_split_sizes_0, x = inputs_13)[name = tensor("op_504")]; + tensor var_506 = sigmoid(x = var_504_1)[name = tensor("op_506")]; + tensor inputs_15 = mul(x = var_504_0, y = var_506)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([5, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_537_begin_0 = const()[name = tensor("op_537_begin_0"), val = tensor([0, -1, 0])]; + tensor var_537_end_0 = const()[name = tensor("op_537_end_0"), val = tensor([5, 16, 256])]; + tensor var_537_end_mask_0 = const()[name = tensor("op_537_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_537 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = conv_out_3)[name = tensor("op_537")]; + tensor var_539_perm_0 = const()[name = tensor("op_539_perm_0"), val = tensor([1, 0, 2])]; + tensor var_539 = transpose(perm = var_539_perm_0, x = var_537)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_539)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_562 = const()[name = tensor("op_562"), val = tensor(0x1p-1)]; + tensor var_563 = mul(x = input_79, y = var_562)[name = tensor("op_563")]; + tensor input_81 = add(x = var_563, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_29, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_592 = const()[name = tensor("op_592"), val = tensor(0x1p-1)]; + tensor var_593 = mul(x = input_91, y = var_592)[name = tensor("op_593")]; + tensor input_93 = add(x = var_593, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_29, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_607 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_608 = const()[name = tensor("op_608"), val = tensor([1, 5, 4, 64])]; + tensor var_609 = reshape(shape = var_608, x = var_607)[name = tensor("op_609")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_613 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_614 = const()[name = tensor("op_614"), val = tensor(0x1p-3)]; + tensor var_615 = mul(x = var_613, y = var_614)[name = tensor("op_615")]; + tensor var_616 = const()[name = tensor("op_616"), val = tensor([1, 5, 4, 64])]; + tensor var_617 = reshape(shape = var_616, x = var_615)[name = tensor("op_617")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_621 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_622 = const()[name = tensor("op_622"), val = tensor([1, 5, 4, 64])]; + tensor var_623 = reshape(shape = var_622, x = var_621)[name = tensor("op_623")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_617)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_609)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_633 = const()[name = tensor("op_633"), val = tensor([5, 1])]; + tensor var_634 = reshape(shape = var_633, x = sqrt_s_t_5)[name = tensor("op_634")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_634)[name = tensor("M_5")]; + tensor var_636 = mul(x = qk_5, y = M_5)[name = tensor("op_636")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_623)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_636, y = v_5)[name = tensor("inner_5")]; + tensor var_638_transpose_x_0 = const()[name = tensor("op_638_transpose_x_0"), val = tensor(false)]; + tensor var_638_transpose_y_0 = const()[name = tensor("op_638_transpose_y_0"), val = tensor(false)]; + tensor var_638 = matmul(transpose_x = var_638_transpose_x_0, transpose_y = var_638_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_638")]; + tensor var_639 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_639")]; + tensor var_640 = const()[name = tensor("op_640"), val = tensor([1, 1, 5, 1])]; + tensor var_641 = reshape(shape = var_640, x = var_639)[name = tensor("op_641")]; + tensor cross_5 = mul(x = var_638, y = var_641)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_644 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_644")]; + tensor var_646_transpose_x_1 = const()[name = tensor("op_646_transpose_x_1"), val = tensor(true)]; + tensor var_646_transpose_y_1 = const()[name = tensor("op_646_transpose_y_1"), val = tensor(false)]; + tensor var_646 = matmul(transpose_x = var_646_transpose_x_1, transpose_y = var_646_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_646")]; + tensor new_kv_unnorm_5 = add(x = var_644, y = var_646)[name = tensor("new_kv_unnorm_5")]; + tensor var_648 = const()[name = tensor("op_648"), val = tensor(0x1.4p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_648)[name = tensor("new_scale_5")]; + tensor var_650 = sqrt(x = new_scale_5)[name = tensor("op_650")]; + tensor var_651 = real_div(x = new_kv_unnorm_5, y = var_650)[name = tensor("op_651")]; + tensor var_652_perm_0 = const()[name = tensor("op_652_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_652 = transpose(perm = var_652_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_18, x = var_652)[name = tensor("out_15")]; + tensor var_656 = const()[name = tensor("op_656"), val = tensor([1, 5, 256])]; + tensor out_17 = reshape(shape = var_656, x = out_15)[name = tensor("out_17")]; + tensor var_658 = silu(x = input_97)[name = tensor("op_658")]; + tensor input_99 = mul(x = var_658, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; + tensor var_666_begin_0 = const()[name = tensor("op_666_begin_0"), val = tensor([0, 0, 0])]; + tensor var_666_end_0 = const()[name = tensor("op_666_end_0"), val = tensor([1, 1, 256])]; + tensor var_666_end_mask_0 = const()[name = tensor("op_666_end_mask_0"), val = tensor([true, false, true])]; + tensor var_666 = slice_by_index(begin = var_666_begin_0, end = var_666_end_0, end_mask = var_666_end_mask_0, x = x_15)[name = tensor("op_666")]; + tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 1, 0])]; + tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 16, 256])]; + tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, true])]; + tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = window_25)[name = tensor("op_669")]; + tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; + tensor window_27 = concat(axis = var_27, interleave = window_27_interleave_0, values = (var_669, var_666))[name = tensor("window_27")]; + tensor var_674_begin_0 = const()[name = tensor("op_674_begin_0"), val = tensor([0, 1, 0])]; + tensor var_674_end_0 = const()[name = tensor("op_674_end_0"), val = tensor([1, 2, 256])]; + tensor var_674_end_mask_0 = const()[name = tensor("op_674_end_mask_0"), val = tensor([true, false, true])]; + tensor var_674 = slice_by_index(begin = var_674_begin_0, end = var_674_end_0, end_mask = var_674_end_mask_0, x = x_15)[name = tensor("op_674")]; + tensor var_677_begin_0 = const()[name = tensor("op_677_begin_0"), val = tensor([0, 1, 0])]; + tensor var_677_end_0 = const()[name = tensor("op_677_end_0"), val = tensor([1, 16, 256])]; + tensor var_677_end_mask_0 = const()[name = tensor("op_677_end_mask_0"), val = tensor([true, true, true])]; + tensor var_677 = slice_by_index(begin = var_677_begin_0, end = var_677_end_0, end_mask = var_677_end_mask_0, x = window_27)[name = tensor("op_677")]; + tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; + tensor window_29 = concat(axis = var_27, interleave = window_29_interleave_0, values = (var_677, var_674))[name = tensor("window_29")]; + tensor var_682_begin_0 = const()[name = tensor("op_682_begin_0"), val = tensor([0, 2, 0])]; + tensor var_682_end_0 = const()[name = tensor("op_682_end_0"), val = tensor([1, 3, 256])]; + tensor var_682_end_mask_0 = const()[name = tensor("op_682_end_mask_0"), val = tensor([true, false, true])]; + tensor var_682 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = x_15)[name = tensor("op_682")]; + tensor var_685_begin_0 = const()[name = tensor("op_685_begin_0"), val = tensor([0, 1, 0])]; + tensor var_685_end_0 = const()[name = tensor("op_685_end_0"), val = tensor([1, 16, 256])]; + tensor var_685_end_mask_0 = const()[name = tensor("op_685_end_mask_0"), val = tensor([true, true, true])]; + tensor var_685 = slice_by_index(begin = var_685_begin_0, end = var_685_end_0, end_mask = var_685_end_mask_0, x = window_29)[name = tensor("op_685")]; + tensor window_31_interleave_0 = const()[name = tensor("window_31_interleave_0"), val = tensor(false)]; + tensor window_31 = concat(axis = var_27, interleave = window_31_interleave_0, values = (var_685, var_682))[name = tensor("window_31")]; + tensor var_690_begin_0 = const()[name = tensor("op_690_begin_0"), val = tensor([0, 3, 0])]; + tensor var_690_end_0 = const()[name = tensor("op_690_end_0"), val = tensor([1, 4, 256])]; + tensor var_690_end_mask_0 = const()[name = tensor("op_690_end_mask_0"), val = tensor([true, false, true])]; + tensor var_690 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, x = x_15)[name = tensor("op_690")]; + tensor var_693_begin_0 = const()[name = tensor("op_693_begin_0"), val = tensor([0, 1, 0])]; + tensor var_693_end_0 = const()[name = tensor("op_693_end_0"), val = tensor([1, 16, 256])]; + tensor var_693_end_mask_0 = const()[name = tensor("op_693_end_mask_0"), val = tensor([true, true, true])]; + tensor var_693 = slice_by_index(begin = var_693_begin_0, end = var_693_end_0, end_mask = var_693_end_mask_0, x = window_31)[name = tensor("op_693")]; + tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; + tensor window_33 = concat(axis = var_27, interleave = window_33_interleave_0, values = (var_693, var_690))[name = tensor("window_33")]; + tensor var_698_begin_0 = const()[name = tensor("op_698_begin_0"), val = tensor([0, 4, 0])]; + tensor var_698_end_0 = const()[name = tensor("op_698_end_0"), val = tensor([1, 1, 256])]; + tensor var_698_end_mask_0 = const()[name = tensor("op_698_end_mask_0"), val = tensor([true, true, true])]; + tensor var_698 = slice_by_index(begin = var_698_begin_0, end = var_698_end_0, end_mask = var_698_end_mask_0, x = x_15)[name = tensor("op_698")]; + tensor var_701_begin_0 = const()[name = tensor("op_701_begin_0"), val = tensor([0, 1, 0])]; + tensor var_701_end_0 = const()[name = tensor("op_701_end_0"), val = tensor([1, 16, 256])]; + tensor var_701_end_mask_0 = const()[name = tensor("op_701_end_mask_0"), val = tensor([true, true, true])]; + tensor var_701 = slice_by_index(begin = var_701_begin_0, end = var_701_end_0, end_mask = var_701_end_mask_0, x = window_33)[name = tensor("op_701")]; + tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; + tensor window_35 = concat(axis = var_27, interleave = window_35_interleave_0, values = (var_701, var_698))[name = tensor("window_35")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_24, interleave = input_101_interleave_0, values = (window_27, window_29, window_31, window_33, window_35))[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_726_split_sizes_0 = const()[name = tensor("op_726_split_sizes_0"), val = tensor([256, 256])]; + tensor var_726_axis_0 = const()[name = tensor("op_726_axis_0"), val = tensor(1)]; + tensor var_726_0, tensor var_726_1 = split(axis = var_726_axis_0, split_sizes = var_726_split_sizes_0, x = inputs_23)[name = tensor("op_726")]; + tensor var_728 = sigmoid(x = var_726_1)[name = tensor("op_728")]; + tensor inputs_25 = mul(x = var_726_0, y = var_728)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([5, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([0, -1, 0])]; + tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([5, 16, 256])]; + tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_759 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, x = conv_out_5)[name = tensor("op_759")]; + tensor var_761_perm_0 = const()[name = tensor("op_761_perm_0"), val = tensor([1, 0, 2])]; + tensor var_761 = transpose(perm = var_761_perm_0, x = var_759)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_761)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_784 = const()[name = tensor("op_784"), val = tensor(0x1p-1)]; + tensor var_785 = mul(x = input_119, y = var_784)[name = tensor("op_785")]; + tensor input_121 = add(x = var_785, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_29, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_814 = const()[name = tensor("op_814"), val = tensor(0x1p-1)]; + tensor var_815 = mul(x = input_131, y = var_814)[name = tensor("op_815")]; + tensor input_133 = add(x = var_815, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_29, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_829 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_830 = const()[name = tensor("op_830"), val = tensor([1, 5, 4, 64])]; + tensor var_831 = reshape(shape = var_830, x = var_829)[name = tensor("op_831")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_835 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_836 = const()[name = tensor("op_836"), val = tensor(0x1p-3)]; + tensor var_837 = mul(x = var_835, y = var_836)[name = tensor("op_837")]; + tensor var_838 = const()[name = tensor("op_838"), val = tensor([1, 5, 4, 64])]; + tensor var_839 = reshape(shape = var_838, x = var_837)[name = tensor("op_839")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_843 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_844 = const()[name = tensor("op_844"), val = tensor([1, 5, 4, 64])]; + tensor var_845 = reshape(shape = var_844, x = var_843)[name = tensor("op_845")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_839)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_831)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_855 = const()[name = tensor("op_855"), val = tensor([5, 1])]; + tensor var_856 = reshape(shape = var_855, x = sqrt_s_t_7)[name = tensor("op_856")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_856)[name = tensor("M_7")]; + tensor var_858 = mul(x = qk_7, y = M_7)[name = tensor("op_858")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_845)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_858, y = v_7)[name = tensor("inner_7")]; + tensor var_860_transpose_x_0 = const()[name = tensor("op_860_transpose_x_0"), val = tensor(false)]; + tensor var_860_transpose_y_0 = const()[name = tensor("op_860_transpose_y_0"), val = tensor(false)]; + tensor var_860 = matmul(transpose_x = var_860_transpose_x_0, transpose_y = var_860_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_860")]; + tensor var_861 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_861")]; + tensor var_862 = const()[name = tensor("op_862"), val = tensor([1, 1, 5, 1])]; + tensor var_863 = reshape(shape = var_862, x = var_861)[name = tensor("op_863")]; + tensor cross_7 = mul(x = var_860, y = var_863)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_866 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_866")]; + tensor var_868_transpose_x_1 = const()[name = tensor("op_868_transpose_x_1"), val = tensor(true)]; + tensor var_868_transpose_y_1 = const()[name = tensor("op_868_transpose_y_1"), val = tensor(false)]; + tensor var_868 = matmul(transpose_x = var_868_transpose_x_1, transpose_y = var_868_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_868")]; + tensor new_kv_unnorm_7 = add(x = var_866, y = var_868)[name = tensor("new_kv_unnorm_7")]; + tensor var_870 = const()[name = tensor("op_870"), val = tensor(0x1.4p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_870)[name = tensor("new_scale_7")]; + tensor var_872 = sqrt(x = new_scale_7)[name = tensor("op_872")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_872)[name = tensor("nkv_1")]; + tensor var_874_perm_0 = const()[name = tensor("op_874_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_874 = transpose(perm = var_874_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_18, x = var_874)[name = tensor("out_21")]; + tensor var_878 = const()[name = tensor("op_878"), val = tensor([1, 5, 256])]; + tensor out_23 = reshape(shape = var_878, x = out_21)[name = tensor("out_23")]; + tensor var_880 = silu(x = input_137)[name = tensor("op_880")]; + tensor input_139 = mul(x = var_880, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_37_begin_0 = const()[name = tensor("window_37_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_37_end_0 = const()[name = tensor("window_37_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_37_end_mask_0 = const()[name = tensor("window_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_37_squeeze_mask_0 = const()[name = tensor("window_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_37 = slice_by_index(begin = window_37_begin_0, end = window_37_end_0, end_mask = window_37_end_mask_0, squeeze_mask = window_37_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_37")]; + tensor var_888_begin_0 = const()[name = tensor("op_888_begin_0"), val = tensor([0, 0, 0])]; + tensor var_888_end_0 = const()[name = tensor("op_888_end_0"), val = tensor([1, 1, 256])]; + tensor var_888_end_mask_0 = const()[name = tensor("op_888_end_mask_0"), val = tensor([true, false, true])]; + tensor var_888 = slice_by_index(begin = var_888_begin_0, end = var_888_end_0, end_mask = var_888_end_mask_0, x = x_21)[name = tensor("op_888")]; + tensor var_891_begin_0 = const()[name = tensor("op_891_begin_0"), val = tensor([0, 1, 0])]; + tensor var_891_end_0 = const()[name = tensor("op_891_end_0"), val = tensor([1, 16, 256])]; + tensor var_891_end_mask_0 = const()[name = tensor("op_891_end_mask_0"), val = tensor([true, true, true])]; + tensor var_891 = slice_by_index(begin = var_891_begin_0, end = var_891_end_0, end_mask = var_891_end_mask_0, x = window_37)[name = tensor("op_891")]; + tensor window_39_interleave_0 = const()[name = tensor("window_39_interleave_0"), val = tensor(false)]; + tensor window_39 = concat(axis = var_27, interleave = window_39_interleave_0, values = (var_891, var_888))[name = tensor("window_39")]; + tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 1, 0])]; + tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 2, 256])]; + tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, false, true])]; + tensor var_896 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = x_21)[name = tensor("op_896")]; + tensor var_899_begin_0 = const()[name = tensor("op_899_begin_0"), val = tensor([0, 1, 0])]; + tensor var_899_end_0 = const()[name = tensor("op_899_end_0"), val = tensor([1, 16, 256])]; + tensor var_899_end_mask_0 = const()[name = tensor("op_899_end_mask_0"), val = tensor([true, true, true])]; + tensor var_899 = slice_by_index(begin = var_899_begin_0, end = var_899_end_0, end_mask = var_899_end_mask_0, x = window_39)[name = tensor("op_899")]; + tensor window_41_interleave_0 = const()[name = tensor("window_41_interleave_0"), val = tensor(false)]; + tensor window_41 = concat(axis = var_27, interleave = window_41_interleave_0, values = (var_899, var_896))[name = tensor("window_41")]; + tensor var_904_begin_0 = const()[name = tensor("op_904_begin_0"), val = tensor([0, 2, 0])]; + tensor var_904_end_0 = const()[name = tensor("op_904_end_0"), val = tensor([1, 3, 256])]; + tensor var_904_end_mask_0 = const()[name = tensor("op_904_end_mask_0"), val = tensor([true, false, true])]; + tensor var_904 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = x_21)[name = tensor("op_904")]; + tensor var_907_begin_0 = const()[name = tensor("op_907_begin_0"), val = tensor([0, 1, 0])]; + tensor var_907_end_0 = const()[name = tensor("op_907_end_0"), val = tensor([1, 16, 256])]; + tensor var_907_end_mask_0 = const()[name = tensor("op_907_end_mask_0"), val = tensor([true, true, true])]; + tensor var_907 = slice_by_index(begin = var_907_begin_0, end = var_907_end_0, end_mask = var_907_end_mask_0, x = window_41)[name = tensor("op_907")]; + tensor window_43_interleave_0 = const()[name = tensor("window_43_interleave_0"), val = tensor(false)]; + tensor window_43 = concat(axis = var_27, interleave = window_43_interleave_0, values = (var_907, var_904))[name = tensor("window_43")]; + tensor var_912_begin_0 = const()[name = tensor("op_912_begin_0"), val = tensor([0, 3, 0])]; + tensor var_912_end_0 = const()[name = tensor("op_912_end_0"), val = tensor([1, 4, 256])]; + tensor var_912_end_mask_0 = const()[name = tensor("op_912_end_mask_0"), val = tensor([true, false, true])]; + tensor var_912 = slice_by_index(begin = var_912_begin_0, end = var_912_end_0, end_mask = var_912_end_mask_0, x = x_21)[name = tensor("op_912")]; + tensor var_915_begin_0 = const()[name = tensor("op_915_begin_0"), val = tensor([0, 1, 0])]; + tensor var_915_end_0 = const()[name = tensor("op_915_end_0"), val = tensor([1, 16, 256])]; + tensor var_915_end_mask_0 = const()[name = tensor("op_915_end_mask_0"), val = tensor([true, true, true])]; + tensor var_915 = slice_by_index(begin = var_915_begin_0, end = var_915_end_0, end_mask = var_915_end_mask_0, x = window_43)[name = tensor("op_915")]; + tensor window_45_interleave_0 = const()[name = tensor("window_45_interleave_0"), val = tensor(false)]; + tensor window_45 = concat(axis = var_27, interleave = window_45_interleave_0, values = (var_915, var_912))[name = tensor("window_45")]; + tensor var_920_begin_0 = const()[name = tensor("op_920_begin_0"), val = tensor([0, 4, 0])]; + tensor var_920_end_0 = const()[name = tensor("op_920_end_0"), val = tensor([1, 1, 256])]; + tensor var_920_end_mask_0 = const()[name = tensor("op_920_end_mask_0"), val = tensor([true, true, true])]; + tensor var_920 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = x_21)[name = tensor("op_920")]; + tensor var_923_begin_0 = const()[name = tensor("op_923_begin_0"), val = tensor([0, 1, 0])]; + tensor var_923_end_0 = const()[name = tensor("op_923_end_0"), val = tensor([1, 16, 256])]; + tensor var_923_end_mask_0 = const()[name = tensor("op_923_end_mask_0"), val = tensor([true, true, true])]; + tensor var_923 = slice_by_index(begin = var_923_begin_0, end = var_923_end_0, end_mask = var_923_end_mask_0, x = window_45)[name = tensor("op_923")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_27, interleave = window_interleave_0, values = (var_923, var_920))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_24, interleave = input_141_interleave_0, values = (window_39, window_41, window_43, window_45, window))[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_948_split_sizes_0 = const()[name = tensor("op_948_split_sizes_0"), val = tensor([256, 256])]; + tensor var_948_axis_0 = const()[name = tensor("op_948_axis_0"), val = tensor(1)]; + tensor var_948_0, tensor var_948_1 = split(axis = var_948_axis_0, split_sizes = var_948_split_sizes_0, x = inputs_33)[name = tensor("op_948")]; + tensor var_950 = sigmoid(x = var_948_1)[name = tensor("op_950")]; + tensor inputs_35 = mul(x = var_948_0, y = var_950)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([5, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_981_begin_0 = const()[name = tensor("op_981_begin_0"), val = tensor([0, -1, 0])]; + tensor var_981_end_0 = const()[name = tensor("op_981_end_0"), val = tensor([5, 16, 256])]; + tensor var_981_end_mask_0 = const()[name = tensor("op_981_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_981 = slice_by_index(begin = var_981_begin_0, end = var_981_end_0, end_mask = var_981_end_mask_0, x = conv_out_7)[name = tensor("op_981")]; + tensor var_983_perm_0 = const()[name = tensor("op_983_perm_0"), val = tensor([1, 0, 2])]; + tensor var_983 = transpose(perm = var_983_perm_0, x = var_981)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_983)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_1006 = const()[name = tensor("op_1006"), val = tensor(0x1p-1)]; + tensor var_1007 = mul(x = input_159, y = var_1006)[name = tensor("op_1007")]; + tensor input_161 = add(x = var_1007, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_29, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_21, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1025_begin_0 = const()[name = tensor("op_1025_begin_0"), val = tensor([0, 0, 5])]; + tensor var_1025_end_0 = const()[name = tensor("op_1025_end_0"), val = tensor([1, 256, 23])]; + tensor var_1025_end_mask_0 = const()[name = tensor("op_1025_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1025_begin_0, end = var_1025_end_0, end_mask = var_1025_end_mask_0, x = cat)[name = tensor("op_1025")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1028 = reduce_l2_norm(axes = var_1027, keep_dims = var_30, x = input_163)[name = tensor("op_1028")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_1028)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_1032_axis_0 = const()[name = tensor("op_1032_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1032_axis_0, values = (var_207, var_429, var_651, nkv_1))[name = tensor("op_1032")]; + tensor var_1034_axis_0 = const()[name = tensor("op_1034_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1034_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1034")]; + tensor var_1036_axis_0 = const()[name = tensor("op_1036_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1036_axis_0, values = (window_11, window_23, window_35, window))[name = tensor("op_1036")]; + tensor var_1045 = const()[name = tensor("op_1045"), val = tensor(0x1.5798eep-27)]; + tensor var_1050 = const()[name = tensor("op_1050"), val = tensor(0x1.4f8b58p-17)]; + tensor var_1052 = const()[name = tensor("op_1052"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_1053 = const()[name = tensor("op_1053"), val = tensor(true)]; + tensor var_1055 = const()[name = tensor("op_1055"), val = tensor(0x1p+0)]; + tensor var_1059 = const()[name = tensor("op_1059"), val = tensor(-1)]; + tensor var_1065 = const()[name = tensor("op_1065"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395712)))]; + tensor var_1127_axes_0 = const()[name = tensor("op_1127_axes_0"), val = tensor([2])]; + tensor var_1127 = expand_dims(axes = var_1127_axes_0, x = emb)[name = tensor("op_1127")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 9, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1127)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_1059, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1135_perm_0 = const()[name = tensor("op_1135_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1139 = const()[name = tensor("op_1139"), val = tensor([9, 5, 256])]; + tensor var_1135 = transpose(perm = var_1135_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1139, x = var_1135)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 9, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1147 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([9, 5, 4, 64])]; + tensor var_1149 = reshape(shape = var_1148, x = var_1147)[name = tensor("op_1149")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1153 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor(0x1p-3)]; + tensor var_1155 = mul(x = var_1153, y = var_1154)[name = tensor("op_1155")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([9, 5, 4, 64])]; + tensor var_1157 = reshape(shape = var_1156, x = var_1155)[name = tensor("op_1157")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1161 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor([9, 5, 4, 64])]; + tensor var_1163 = reshape(shape = var_1162, x = var_1161)[name = tensor("op_1163")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_1065, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_1055, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1157)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1149)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 5])]; + tensor var_1176 = reshape(shape = var_1175, x = valid_mask)[name = tensor("op_1176")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1176)[name = tensor("causal_with_valid_1")]; + tensor var_1178 = const()[name = tensor("op_1178"), val = tensor([5, 1])]; + tensor var_1179 = reshape(shape = var_1178, x = sqrt_s_t_9)[name = tensor("op_1179")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1179)[name = tensor("M_9")]; + tensor var_1181 = mul(x = qk_9, y = M_9)[name = tensor("op_1181")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1163)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1181, y = v_9)[name = tensor("inner_9")]; + tensor var_1183_transpose_x_0 = const()[name = tensor("op_1183_transpose_x_0"), val = tensor(false)]; + tensor var_1183_transpose_y_0 = const()[name = tensor("op_1183_transpose_y_0"), val = tensor(false)]; + tensor var_1183 = matmul(transpose_x = var_1183_transpose_x_0, transpose_y = var_1183_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1183")]; + tensor var_1184 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1184")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 5, 1])]; + tensor var_1186 = reshape(shape = var_1185, x = var_1184)[name = tensor("op_1186")]; + tensor cross_9 = mul(x = var_1183, y = var_1186)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1, 1, 5, 1])]; + tensor var_1190 = reshape(shape = var_1189, x = valid_mask)[name = tensor("op_1190")]; + tensor v_masked_1 = mul(x = v_9, y = var_1190)[name = tensor("v_masked_1")]; + tensor var_1192 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1192")]; + tensor var_1194_transpose_x_1 = const()[name = tensor("op_1194_transpose_x_1"), val = tensor(true)]; + tensor var_1194_transpose_y_1 = const()[name = tensor("op_1194_transpose_y_1"), val = tensor(false)]; + tensor var_1194 = matmul(transpose_x = var_1194_transpose_x_1, transpose_y = var_1194_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1194")]; + tensor new_kv_unnorm_9 = add(x = var_1192, y = var_1194)[name = tensor("new_kv_unnorm_9")]; + tensor var_1196_keep_dims_0 = const()[name = tensor("op_1196_keep_dims_0"), val = tensor(false)]; + tensor var_1196 = reduce_sum(keep_dims = var_1196_keep_dims_0, x = valid_mask)[name = tensor("op_1196")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1])]; + tensor var_1198 = reshape(shape = var_1197, x = var_1196)[name = tensor("op_1198")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1198)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_1055, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1202 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1202")]; + tensor var_1203_perm_0 = const()[name = tensor("op_1203_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1203 = transpose(perm = var_1203_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1052, x = var_1203)[name = tensor("out_27")]; + tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([9, 5, 256])]; + tensor out_29 = reshape(shape = var_1207, x = out_27)[name = tensor("out_29")]; + tensor var_1209 = silu(x = input_169)[name = tensor("op_1209")]; + tensor input_171 = mul(x = var_1209, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1050, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1219 = const()[name = tensor("op_1219"), val = tensor([1, 9, 5, 256])]; + tensor var_1220 = reshape(shape = var_1219, x = xt_1)[name = tensor("op_1220")]; + tensor var_1221_perm_0 = const()[name = tensor("op_1221_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([5, 9, 256])]; + tensor var_1221 = transpose(perm = var_1221_perm_0, x = var_1220)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1224, x = var_1221)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1247 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([9, 5, 3, 256])]; + tensor var_1249 = reshape(shape = concat_1, x = var_1247)[name = tensor("op_1249")]; + tensor var_1250_axes_0 = const()[name = tensor("op_1250_axes_0"), val = tensor([0])]; + tensor var_1250 = expand_dims(axes = var_1250_axes_0, x = var_1249)[name = tensor("op_1250")]; + tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1252_axes_0 = const()[name = tensor("op_1252_axes_0"), val = tensor([-2])]; + tensor var_1251 = transpose(perm = var_1251_perm_0, x = var_1250)[name = tensor("transpose_21")]; + tensor var_1252 = squeeze(axes = var_1252_axes_0, x = var_1251)[name = tensor("op_1252")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 9, 5, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1252)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 9, 5, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1252)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 9, 5, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1252)[name = tensor("v_11")]; + tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([9, 20, 64])]; + tensor var_1261 = reshape(shape = var_1260, x = q_11)[name = tensor("op_1261")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1267 = const()[name = tensor("op_1267"), val = tensor([9, 20, 64])]; + tensor var_1268 = reshape(shape = var_1267, x = k_11)[name = tensor("op_1268")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([9, 20, 64])]; + tensor var_1275 = reshape(shape = var_1274, x = v_11)[name = tensor("op_1275")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([5, 4, 9, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1261)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1278, x = q_13)[name = tensor("q_15")]; + tensor var_1280 = const()[name = tensor("op_1280"), val = tensor([5, 4, 9, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1268)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1280, x = k_13)[name = tensor("k_15")]; + tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([5, 4, 9, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1275)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1282, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1285 = const()[name = tensor("op_1285"), val = tensor([2, 0, 1, 3])]; + tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([45, 256])]; + tensor var_1286 = transpose(perm = var_1285, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1290, x = var_1286)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([9, 5, 256])]; + tensor attn_output_7 = reshape(shape = var_1294, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1050, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1050, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 5, 9, 256])]; + tensor x_31 = reshape(shape = var_1314, x = xt_3)[name = tensor("x_31")]; + tensor var_1316_perm_0 = const()[name = tensor("op_1316_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1320 = const()[name = tensor("op_1320"), val = tensor([9, 5, 256])]; + tensor var_1316 = transpose(perm = var_1316_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1320, x = var_1316)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 9, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1328 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([9, 5, 4, 64])]; + tensor var_1330 = reshape(shape = var_1329, x = var_1328)[name = tensor("op_1330")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1334 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor(0x1p-3)]; + tensor var_1336 = mul(x = var_1334, y = var_1335)[name = tensor("op_1336")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([9, 5, 4, 64])]; + tensor var_1338 = reshape(shape = var_1337, x = var_1336)[name = tensor("op_1338")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1342 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([9, 5, 4, 64])]; + tensor var_1344 = reshape(shape = var_1343, x = var_1342)[name = tensor("op_1344")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_1055, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1338)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1330)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1359 = const()[name = tensor("op_1359"), val = tensor([5, 1])]; + tensor var_1360 = reshape(shape = var_1359, x = sqrt_s_t)[name = tensor("op_1360")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1360)[name = tensor("M")]; + tensor var_1362 = mul(x = qk, y = M)[name = tensor("op_1362")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1344)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1362, y = v_17)[name = tensor("inner")]; + tensor var_1364_transpose_x_0 = const()[name = tensor("op_1364_transpose_x_0"), val = tensor(false)]; + tensor var_1364_transpose_y_0 = const()[name = tensor("op_1364_transpose_y_0"), val = tensor(false)]; + tensor var_1364 = matmul(transpose_x = var_1364_transpose_x_0, transpose_y = var_1364_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1364")]; + tensor var_1365 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1365")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 5, 1])]; + tensor var_1367 = reshape(shape = var_1366, x = var_1365)[name = tensor("op_1367")]; + tensor cross = mul(x = var_1364, y = var_1367)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1190)[name = tensor("v_masked")]; + tensor var_1373 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1373")]; + tensor var_1375_transpose_x_1 = const()[name = tensor("op_1375_transpose_x_1"), val = tensor(true)]; + tensor var_1375_transpose_y_1 = const()[name = tensor("op_1375_transpose_y_1"), val = tensor(false)]; + tensor var_1375 = matmul(transpose_x = var_1375_transpose_x_1, transpose_y = var_1375_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1375")]; + tensor new_kv_unnorm = add(x = var_1373, y = var_1375)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1198)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_1055, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1384_perm_0 = const()[name = tensor("op_1384_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1384 = transpose(perm = var_1384_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1052, x = var_1384)[name = tensor("out_33")]; + tensor var_1388 = const()[name = tensor("op_1388"), val = tensor([9, 5, 256])]; + tensor out = reshape(shape = var_1388, x = out_33)[name = tensor("out")]; + tensor var_1390 = silu(x = input_187)[name = tensor("op_1390")]; + tensor input_189 = mul(x = var_1390, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1050, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 9, 5, 256])]; + tensor var_1401 = reshape(shape = var_1400, x = xt_5)[name = tensor("op_1401")]; + tensor var_1402_perm_0 = const()[name = tensor("op_1402_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1405 = const()[name = tensor("op_1405"), val = tensor([5, 9, 256])]; + tensor var_1402 = transpose(perm = var_1402_perm_0, x = var_1401)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1405, x = var_1402)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1428 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([9, 5, 3, 256])]; + tensor var_1430 = reshape(shape = concat_2, x = var_1428)[name = tensor("op_1430")]; + tensor var_1431_axes_0 = const()[name = tensor("op_1431_axes_0"), val = tensor([0])]; + tensor var_1431 = expand_dims(axes = var_1431_axes_0, x = var_1430)[name = tensor("op_1431")]; + tensor var_1432_perm_0 = const()[name = tensor("op_1432_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1433_axes_0 = const()[name = tensor("op_1433_axes_0"), val = tensor([-2])]; + tensor var_1432 = transpose(perm = var_1432_perm_0, x = var_1431)[name = tensor("transpose_8")]; + tensor var_1433 = squeeze(axes = var_1433_axes_0, x = var_1432)[name = tensor("op_1433")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 9, 5, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1433)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 9, 5, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1433)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 9, 5, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1433)[name = tensor("v_19")]; + tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([9, 20, 64])]; + tensor var_1442 = reshape(shape = var_1441, x = q_19)[name = tensor("op_1442")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([9, 20, 64])]; + tensor var_1449 = reshape(shape = var_1448, x = k_19)[name = tensor("op_1449")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([9, 20, 64])]; + tensor var_1456 = reshape(shape = var_1455, x = v_19)[name = tensor("op_1456")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([5, 4, 9, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1442)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1459, x = q_21)[name = tensor("q")]; + tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([5, 4, 9, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1449)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1461, x = k_21)[name = tensor("k")]; + tensor var_1463 = const()[name = tensor("op_1463"), val = tensor([5, 4, 9, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1456)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1463, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1466 = const()[name = tensor("op_1466"), val = tensor([2, 0, 1, 3])]; + tensor var_1471 = const()[name = tensor("op_1471"), val = tensor([45, 256])]; + tensor var_1467 = transpose(perm = var_1466, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1471, x = var_1467)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1475 = const()[name = tensor("op_1475"), val = tensor([9, 5, 256])]; + tensor attn_output = reshape(shape = var_1475, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1050, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1050, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1495 = const()[name = tensor("op_1495"), val = tensor([1, 5, 9, 256])]; + tensor input = reshape(shape = var_1495, x = xt)[name = tensor("input")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([-1])]; + tensor var_1498 = reduce_l2_norm(axes = var_1497, keep_dims = var_1053, x = input)[name = tensor("op_1498")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_1045, beta = const_42, x = var_1498)[name = tensor("clip_5")]; + tensor var_1500 = real_div(x = input, y = clip_5)[name = tensor("op_1500")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([5, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([5, 256, 9])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1500)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 5, 9])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 5, 8])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1504")]; + tensor var_1506_axis_0 = const()[name = tensor("op_1506_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1506_axis_0, values = (var_1202, nkv))[name = tensor("op_1506")]; + tensor var_1508_axis_0 = const()[name = tensor("op_1508_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1508_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1508")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/weights/weight.bin b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..a06142e2373f884c2f7932f9c13386391a7b30f2 --- /dev/null +++ b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81e0a5761f683bb2049d3d114ed8b416833a8b18a0b2585518f3496c9c0283ed +size 44441856 diff --git a/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..e15cb45b1297e6fc6556ae5ba31a2b881b5eb03b --- /dev/null +++ b/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32b8d373605b82b210571b934838ef9e0ede35dc97c852d616acea69d2d1c22c +size 196620 diff --git a/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..a06142e2373f884c2f7932f9c13386391a7b30f2 --- /dev/null +++ b/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81e0a5761f683bb2049d3d114ed8b416833a8b18a0b2585518f3496c9c0283ed +size 44441856 diff --git a/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Manifest.json b/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..7b1a12bb9811d9e20f8485718da28926f0fe3cd1 --- /dev/null +++ b/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "B3E5EFE2-8BD2-41C3-B202-D931C2CCA18E": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "E2319A52-06A1-4D52-8B9C-D3854ED1D555": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "E2319A52-06A1-4D52-8B9C-D3854ED1D555" +} diff --git a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/analytics/coremldata.bin b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..e6a49061db71bd6609e70b767cc7b7d63298f84f --- /dev/null +++ b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4b9f84d7ea9b63709ef0b529a4154d7c460335b9b0ff8de63b48ab92f0e2e2c +size 243 diff --git a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/coremldata.bin b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..2f3d2e94407edd8a732a883b6a075a76fb034d65 --- /dev/null +++ b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b761b596de7a3abcc362d422df9b4f2626ec351b373ef5652bc478cea8e6568 +size 1308 diff --git a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/metadata.json b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..36071321ca497cfeb4a01fcebd77a489dc4f5a14 --- /dev/null +++ b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND DIHARD II streaming diarizer (pipeline, T=1, max_speakers=10)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 10)", + "shortDescription" : "", + "shape" : "[1, 1, 10]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 66, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 36, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 10, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 46 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 345)", + "shortDescription" : "", + "shape" : "[1, 1, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"dih2\", \"model_label\": \"DIHARD II\", \"variant\": \"pipeline\", \"chunk_size\": 1, \"step_duration_ms\": 100, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_dih2_100ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/model.mil b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..182f88767b6a7156fba8d5e17cc71543704b1b0a --- /dev/null +++ b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/model.mil @@ -0,0 +1,1184 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0]])]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 1, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 1, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 1, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 1, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+0)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 1, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, x_3))[name = tensor("window_3")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = window_3)[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_249_split_sizes_0 = const()[name = tensor("op_249_split_sizes_0"), val = tensor([256, 256])]; + tensor var_249_axis_0 = const()[name = tensor("op_249_axis_0"), val = tensor(1)]; + tensor var_249_0, tensor var_249_1 = split(axis = var_249_axis_0, split_sizes = var_249_split_sizes_0, x = inputs_3)[name = tensor("op_249")]; + tensor var_251 = sigmoid(x = var_249_1)[name = tensor("op_251")]; + tensor inputs_5 = mul(x = var_249_0, y = var_251)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([1, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_282_begin_0 = const()[name = tensor("op_282_begin_0"), val = tensor([0, -1, 0])]; + tensor var_282_end_0 = const()[name = tensor("op_282_end_0"), val = tensor([1, 16, 256])]; + tensor var_282_end_mask_0 = const()[name = tensor("op_282_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_282 = slice_by_index(begin = var_282_begin_0, end = var_282_end_0, end_mask = var_282_end_mask_0, x = conv_out_1)[name = tensor("op_282")]; + tensor var_284_perm_0 = const()[name = tensor("op_284_perm_0"), val = tensor([1, 0, 2])]; + tensor var_284 = transpose(perm = var_284_perm_0, x = var_282)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_284)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_307 = const()[name = tensor("op_307"), val = tensor(0x1p-1)]; + tensor var_308 = mul(x = input_39, y = var_307)[name = tensor("op_308")]; + tensor input_41 = add(x = var_308, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_337 = const()[name = tensor("op_337"), val = tensor(0x1p-1)]; + tensor var_338 = mul(x = input_51, y = var_337)[name = tensor("op_338")]; + tensor input_53 = add(x = var_338, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_352 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_353 = const()[name = tensor("op_353"), val = tensor([1, 1, 4, 64])]; + tensor var_354 = reshape(shape = var_353, x = var_352)[name = tensor("op_354")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_358 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_359 = const()[name = tensor("op_359"), val = tensor(0x1p-3)]; + tensor var_360 = mul(x = var_358, y = var_359)[name = tensor("op_360")]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 1, 4, 64])]; + tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_366 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_367 = const()[name = tensor("op_367"), val = tensor([1, 1, 4, 64])]; + tensor var_368 = reshape(shape = var_367, x = var_366)[name = tensor("op_368")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_362)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_354)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 1])]; + tensor var_379 = reshape(shape = var_378, x = sqrt_s_t_3)[name = tensor("op_379")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_379)[name = tensor("M_3")]; + tensor var_381 = mul(x = qk_3, y = M_3)[name = tensor("op_381")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_368)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_381, y = v_3)[name = tensor("inner_3")]; + tensor var_383_transpose_x_0 = const()[name = tensor("op_383_transpose_x_0"), val = tensor(false)]; + tensor var_383_transpose_y_0 = const()[name = tensor("op_383_transpose_y_0"), val = tensor(false)]; + tensor var_383 = matmul(transpose_x = var_383_transpose_x_0, transpose_y = var_383_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_383")]; + tensor var_384 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_384")]; + tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 1, 1, 1])]; + tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; + tensor cross_3 = mul(x = var_383, y = var_386)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_389 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_389")]; + tensor var_391_transpose_x_1 = const()[name = tensor("op_391_transpose_x_1"), val = tensor(true)]; + tensor var_391_transpose_y_1 = const()[name = tensor("op_391_transpose_y_1"), val = tensor(false)]; + tensor var_391 = matmul(transpose_x = var_391_transpose_x_1, transpose_y = var_391_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_391")]; + tensor new_kv_unnorm_3 = add(x = var_389, y = var_391)[name = tensor("new_kv_unnorm_3")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor(0x1p+0)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_393)[name = tensor("new_scale_3")]; + tensor var_395 = sqrt(x = new_scale_3)[name = tensor("op_395")]; + tensor var_396 = real_div(x = new_kv_unnorm_3, y = var_395)[name = tensor("op_396")]; + tensor var_397_perm_0 = const()[name = tensor("op_397_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_397 = transpose(perm = var_397_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_397)[name = tensor("out_9")]; + tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 256])]; + tensor out_11 = reshape(shape = var_401, x = out_9)[name = tensor("out_11")]; + tensor var_403 = silu(x = input_57)[name = tensor("op_403")]; + tensor input_59 = mul(x = var_403, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_5_begin_0 = const()[name = tensor("window_5_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_5_end_0 = const()[name = tensor("window_5_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_5_end_mask_0 = const()[name = tensor("window_5_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_5_squeeze_mask_0 = const()[name = tensor("window_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_5 = slice_by_index(begin = window_5_begin_0, end = window_5_end_0, end_mask = window_5_end_mask_0, squeeze_mask = window_5_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_5")]; + tensor var_414_begin_0 = const()[name = tensor("op_414_begin_0"), val = tensor([0, 1, 0])]; + tensor var_414_end_0 = const()[name = tensor("op_414_end_0"), val = tensor([1, 16, 256])]; + tensor var_414_end_mask_0 = const()[name = tensor("op_414_end_mask_0"), val = tensor([true, true, true])]; + tensor var_414 = slice_by_index(begin = var_414_begin_0, end = var_414_end_0, end_mask = var_414_end_mask_0, x = window_5)[name = tensor("op_414")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_414, x_9))[name = tensor("window_7")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = window_7)[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_439_split_sizes_0 = const()[name = tensor("op_439_split_sizes_0"), val = tensor([256, 256])]; + tensor var_439_axis_0 = const()[name = tensor("op_439_axis_0"), val = tensor(1)]; + tensor var_439_0, tensor var_439_1 = split(axis = var_439_axis_0, split_sizes = var_439_split_sizes_0, x = inputs_13)[name = tensor("op_439")]; + tensor var_441 = sigmoid(x = var_439_1)[name = tensor("op_441")]; + tensor inputs_15 = mul(x = var_439_0, y = var_441)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([1, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_472_begin_0 = const()[name = tensor("op_472_begin_0"), val = tensor([0, -1, 0])]; + tensor var_472_end_0 = const()[name = tensor("op_472_end_0"), val = tensor([1, 16, 256])]; + tensor var_472_end_mask_0 = const()[name = tensor("op_472_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_472 = slice_by_index(begin = var_472_begin_0, end = var_472_end_0, end_mask = var_472_end_mask_0, x = conv_out_3)[name = tensor("op_472")]; + tensor var_474_perm_0 = const()[name = tensor("op_474_perm_0"), val = tensor([1, 0, 2])]; + tensor var_474 = transpose(perm = var_474_perm_0, x = var_472)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_474)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_497 = const()[name = tensor("op_497"), val = tensor(0x1p-1)]; + tensor var_498 = mul(x = input_79, y = var_497)[name = tensor("op_498")]; + tensor input_81 = add(x = var_498, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_527 = const()[name = tensor("op_527"), val = tensor(0x1p-1)]; + tensor var_528 = mul(x = input_91, y = var_527)[name = tensor("op_528")]; + tensor input_93 = add(x = var_528, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_542 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 1, 4, 64])]; + tensor var_544 = reshape(shape = var_543, x = var_542)[name = tensor("op_544")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_548 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_549 = const()[name = tensor("op_549"), val = tensor(0x1p-3)]; + tensor var_550 = mul(x = var_548, y = var_549)[name = tensor("op_550")]; + tensor var_551 = const()[name = tensor("op_551"), val = tensor([1, 1, 4, 64])]; + tensor var_552 = reshape(shape = var_551, x = var_550)[name = tensor("op_552")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_556 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_557 = const()[name = tensor("op_557"), val = tensor([1, 1, 4, 64])]; + tensor var_558 = reshape(shape = var_557, x = var_556)[name = tensor("op_558")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_552)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_544)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_568 = const()[name = tensor("op_568"), val = tensor([1, 1])]; + tensor var_569 = reshape(shape = var_568, x = sqrt_s_t_5)[name = tensor("op_569")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_569)[name = tensor("M_5")]; + tensor var_571 = mul(x = qk_5, y = M_5)[name = tensor("op_571")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_558)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_571, y = v_5)[name = tensor("inner_5")]; + tensor var_573_transpose_x_0 = const()[name = tensor("op_573_transpose_x_0"), val = tensor(false)]; + tensor var_573_transpose_y_0 = const()[name = tensor("op_573_transpose_y_0"), val = tensor(false)]; + tensor var_573 = matmul(transpose_x = var_573_transpose_x_0, transpose_y = var_573_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_573")]; + tensor var_574 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_574")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 1, 1, 1])]; + tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; + tensor cross_5 = mul(x = var_573, y = var_576)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_579 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_579")]; + tensor var_581_transpose_x_1 = const()[name = tensor("op_581_transpose_x_1"), val = tensor(true)]; + tensor var_581_transpose_y_1 = const()[name = tensor("op_581_transpose_y_1"), val = tensor(false)]; + tensor var_581 = matmul(transpose_x = var_581_transpose_x_1, transpose_y = var_581_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_581")]; + tensor new_kv_unnorm_5 = add(x = var_579, y = var_581)[name = tensor("new_kv_unnorm_5")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor(0x1p+0)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_583)[name = tensor("new_scale_5")]; + tensor var_585 = sqrt(x = new_scale_5)[name = tensor("op_585")]; + tensor var_586 = real_div(x = new_kv_unnorm_5, y = var_585)[name = tensor("op_586")]; + tensor var_587_perm_0 = const()[name = tensor("op_587_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_587 = transpose(perm = var_587_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_587)[name = tensor("out_15")]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 256])]; + tensor out_17 = reshape(shape = var_591, x = out_15)[name = tensor("out_17")]; + tensor var_593 = silu(x = input_97)[name = tensor("op_593")]; + tensor input_99 = mul(x = var_593, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; + tensor var_604_begin_0 = const()[name = tensor("op_604_begin_0"), val = tensor([0, 1, 0])]; + tensor var_604_end_0 = const()[name = tensor("op_604_end_0"), val = tensor([1, 16, 256])]; + tensor var_604_end_mask_0 = const()[name = tensor("op_604_end_mask_0"), val = tensor([true, true, true])]; + tensor var_604 = slice_by_index(begin = var_604_begin_0, end = var_604_end_0, end_mask = var_604_end_mask_0, x = window_9)[name = tensor("op_604")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_604, x_15))[name = tensor("window_11")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = window_11)[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_629_split_sizes_0 = const()[name = tensor("op_629_split_sizes_0"), val = tensor([256, 256])]; + tensor var_629_axis_0 = const()[name = tensor("op_629_axis_0"), val = tensor(1)]; + tensor var_629_0, tensor var_629_1 = split(axis = var_629_axis_0, split_sizes = var_629_split_sizes_0, x = inputs_23)[name = tensor("op_629")]; + tensor var_631 = sigmoid(x = var_629_1)[name = tensor("op_631")]; + tensor inputs_25 = mul(x = var_629_0, y = var_631)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([1, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_662_begin_0 = const()[name = tensor("op_662_begin_0"), val = tensor([0, -1, 0])]; + tensor var_662_end_0 = const()[name = tensor("op_662_end_0"), val = tensor([1, 16, 256])]; + tensor var_662_end_mask_0 = const()[name = tensor("op_662_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_662 = slice_by_index(begin = var_662_begin_0, end = var_662_end_0, end_mask = var_662_end_mask_0, x = conv_out_5)[name = tensor("op_662")]; + tensor var_664_perm_0 = const()[name = tensor("op_664_perm_0"), val = tensor([1, 0, 2])]; + tensor var_664 = transpose(perm = var_664_perm_0, x = var_662)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_664)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_687 = const()[name = tensor("op_687"), val = tensor(0x1p-1)]; + tensor var_688 = mul(x = input_119, y = var_687)[name = tensor("op_688")]; + tensor input_121 = add(x = var_688, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_717 = const()[name = tensor("op_717"), val = tensor(0x1p-1)]; + tensor var_718 = mul(x = input_131, y = var_717)[name = tensor("op_718")]; + tensor input_133 = add(x = var_718, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_732 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_733 = const()[name = tensor("op_733"), val = tensor([1, 1, 4, 64])]; + tensor var_734 = reshape(shape = var_733, x = var_732)[name = tensor("op_734")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_738 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_739 = const()[name = tensor("op_739"), val = tensor(0x1p-3)]; + tensor var_740 = mul(x = var_738, y = var_739)[name = tensor("op_740")]; + tensor var_741 = const()[name = tensor("op_741"), val = tensor([1, 1, 4, 64])]; + tensor var_742 = reshape(shape = var_741, x = var_740)[name = tensor("op_742")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_746 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 1, 4, 64])]; + tensor var_748 = reshape(shape = var_747, x = var_746)[name = tensor("op_748")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_742)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_734)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_758 = const()[name = tensor("op_758"), val = tensor([1, 1])]; + tensor var_759 = reshape(shape = var_758, x = sqrt_s_t_7)[name = tensor("op_759")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_759)[name = tensor("M_7")]; + tensor var_761 = mul(x = qk_7, y = M_7)[name = tensor("op_761")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_748)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_761, y = v_7)[name = tensor("inner_7")]; + tensor var_763_transpose_x_0 = const()[name = tensor("op_763_transpose_x_0"), val = tensor(false)]; + tensor var_763_transpose_y_0 = const()[name = tensor("op_763_transpose_y_0"), val = tensor(false)]; + tensor var_763 = matmul(transpose_x = var_763_transpose_x_0, transpose_y = var_763_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_763")]; + tensor var_764 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_764")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 1, 1, 1])]; + tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; + tensor cross_7 = mul(x = var_763, y = var_766)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_769 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_769")]; + tensor var_771_transpose_x_1 = const()[name = tensor("op_771_transpose_x_1"), val = tensor(true)]; + tensor var_771_transpose_y_1 = const()[name = tensor("op_771_transpose_y_1"), val = tensor(false)]; + tensor var_771 = matmul(transpose_x = var_771_transpose_x_1, transpose_y = var_771_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_771")]; + tensor new_kv_unnorm_7 = add(x = var_769, y = var_771)[name = tensor("new_kv_unnorm_7")]; + tensor var_773 = const()[name = tensor("op_773"), val = tensor(0x1p+0)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_773)[name = tensor("new_scale_7")]; + tensor var_775 = sqrt(x = new_scale_7)[name = tensor("op_775")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_775)[name = tensor("nkv_1")]; + tensor var_777_perm_0 = const()[name = tensor("op_777_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_777 = transpose(perm = var_777_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_777)[name = tensor("out_21")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 1, 256])]; + tensor out_23 = reshape(shape = var_781, x = out_21)[name = tensor("out_23")]; + tensor var_783 = silu(x = input_137)[name = tensor("op_783")]; + tensor input_139 = mul(x = var_783, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; + tensor var_794_begin_0 = const()[name = tensor("op_794_begin_0"), val = tensor([0, 1, 0])]; + tensor var_794_end_0 = const()[name = tensor("op_794_end_0"), val = tensor([1, 16, 256])]; + tensor var_794_end_mask_0 = const()[name = tensor("op_794_end_mask_0"), val = tensor([true, true, true])]; + tensor var_794 = slice_by_index(begin = var_794_begin_0, end = var_794_end_0, end_mask = var_794_end_mask_0, x = window_13)[name = tensor("op_794")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_794, x_21))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = window)[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_819_split_sizes_0 = const()[name = tensor("op_819_split_sizes_0"), val = tensor([256, 256])]; + tensor var_819_axis_0 = const()[name = tensor("op_819_axis_0"), val = tensor(1)]; + tensor var_819_0, tensor var_819_1 = split(axis = var_819_axis_0, split_sizes = var_819_split_sizes_0, x = inputs_33)[name = tensor("op_819")]; + tensor var_821 = sigmoid(x = var_819_1)[name = tensor("op_821")]; + tensor inputs_35 = mul(x = var_819_0, y = var_821)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([1, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_852_begin_0 = const()[name = tensor("op_852_begin_0"), val = tensor([0, -1, 0])]; + tensor var_852_end_0 = const()[name = tensor("op_852_end_0"), val = tensor([1, 16, 256])]; + tensor var_852_end_mask_0 = const()[name = tensor("op_852_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_852 = slice_by_index(begin = var_852_begin_0, end = var_852_end_0, end_mask = var_852_end_mask_0, x = conv_out_7)[name = tensor("op_852")]; + tensor var_854_perm_0 = const()[name = tensor("op_854_perm_0"), val = tensor([1, 0, 2])]; + tensor var_854 = transpose(perm = var_854_perm_0, x = var_852)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_854)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_877 = const()[name = tensor("op_877"), val = tensor(0x1p-1)]; + tensor var_878 = mul(x = input_159, y = var_877)[name = tensor("op_878")]; + tensor input_161 = add(x = var_878, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 0, 1])]; + tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 256, 19])]; + tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = cat)[name = tensor("op_896")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_898 = const()[name = tensor("op_898"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_899 = reduce_l2_norm(axes = var_898, keep_dims = var_29, x = input_163)[name = tensor("op_899")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_899)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_903_axis_0 = const()[name = tensor("op_903_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_903_axis_0, values = (var_206, var_396, var_586, nkv_1))[name = tensor("op_903")]; + tensor var_905_axis_0 = const()[name = tensor("op_905_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_905_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_905")]; + tensor var_907_axis_0 = const()[name = tensor("op_907_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_907_axis_0, values = (window_3, window_7, window_11, window))[name = tensor("op_907")]; + tensor var_916 = const()[name = tensor("op_916"), val = tensor(0x1.5798eep-27)]; + tensor var_921 = const()[name = tensor("op_921"), val = tensor(0x1.4f8b58p-17)]; + tensor var_923 = const()[name = tensor("op_923"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_924 = const()[name = tensor("op_924"), val = tensor(true)]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor(0x1p+0)]; + tensor var_930 = const()[name = tensor("op_930"), val = tensor(-1)]; + tensor var_936 = const()[name = tensor("op_936"), val = tensor(0)]; + tensor var_993 = const()[name = tensor("op_993"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_998_axes_0 = const()[name = tensor("op_998_axes_0"), val = tensor([2])]; + tensor var_998 = expand_dims(axes = var_998_axes_0, x = emb)[name = tensor("op_998")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_998)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_930, interleave = input_165_interleave_0, values = (emb_exp, var_993))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1006_perm_0 = const()[name = tensor("op_1006_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([12, 1, 256])]; + tensor var_1006 = transpose(perm = var_1006_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1010, x = var_1006)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1018 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1019 = const()[name = tensor("op_1019"), val = tensor([12, 1, 4, 64])]; + tensor var_1020 = reshape(shape = var_1019, x = var_1018)[name = tensor("op_1020")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1024 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1025 = const()[name = tensor("op_1025"), val = tensor(0x1p-3)]; + tensor var_1026 = mul(x = var_1024, y = var_1025)[name = tensor("op_1026")]; + tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([12, 1, 4, 64])]; + tensor var_1028 = reshape(shape = var_1027, x = var_1026)[name = tensor("op_1028")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1032 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1033 = const()[name = tensor("op_1033"), val = tensor([12, 1, 4, 64])]; + tensor var_1034 = reshape(shape = var_1033, x = var_1032)[name = tensor("op_1034")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_936, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_926, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1028)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1020)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 1])]; + tensor var_1047 = reshape(shape = var_1046, x = valid_mask)[name = tensor("op_1047")]; + tensor var_1049 = const()[name = tensor("op_1049"), val = tensor([1, 1])]; + tensor var_1050 = reshape(shape = var_1049, x = sqrt_s_t_9)[name = tensor("op_1050")]; + tensor M_9 = real_div(x = var_1047, y = var_1050)[name = tensor("M_9")]; + tensor var_1052 = mul(x = qk_9, y = M_9)[name = tensor("op_1052")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1034)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1052, y = v_9)[name = tensor("inner_9")]; + tensor var_1054_transpose_x_0 = const()[name = tensor("op_1054_transpose_x_0"), val = tensor(false)]; + tensor var_1054_transpose_y_0 = const()[name = tensor("op_1054_transpose_y_0"), val = tensor(false)]; + tensor var_1054 = matmul(transpose_x = var_1054_transpose_x_0, transpose_y = var_1054_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1054")]; + tensor var_1055 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1055")]; + tensor var_1056 = const()[name = tensor("op_1056"), val = tensor([1, 1, 1, 1])]; + tensor var_1057 = reshape(shape = var_1056, x = var_1055)[name = tensor("op_1057")]; + tensor cross_9 = mul(x = var_1054, y = var_1057)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1060 = const()[name = tensor("op_1060"), val = tensor([1, 1, 1, 1])]; + tensor var_1061 = reshape(shape = var_1060, x = valid_mask)[name = tensor("op_1061")]; + tensor v_masked_1 = mul(x = v_9, y = var_1061)[name = tensor("v_masked_1")]; + tensor var_1063 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1063")]; + tensor var_1065_transpose_x_1 = const()[name = tensor("op_1065_transpose_x_1"), val = tensor(true)]; + tensor var_1065_transpose_y_1 = const()[name = tensor("op_1065_transpose_y_1"), val = tensor(false)]; + tensor var_1065 = matmul(transpose_x = var_1065_transpose_x_1, transpose_y = var_1065_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1065")]; + tensor new_kv_unnorm_9 = add(x = var_1063, y = var_1065)[name = tensor("new_kv_unnorm_9")]; + tensor var_1067_keep_dims_0 = const()[name = tensor("op_1067_keep_dims_0"), val = tensor(false)]; + tensor var_1067 = reduce_sum(keep_dims = var_1067_keep_dims_0, x = valid_mask)[name = tensor("op_1067")]; + tensor var_1068 = const()[name = tensor("op_1068"), val = tensor([1])]; + tensor var_1069 = reshape(shape = var_1068, x = var_1067)[name = tensor("op_1069")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1069)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_926, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1073 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1073")]; + tensor var_1074_perm_0 = const()[name = tensor("op_1074_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1074 = transpose(perm = var_1074_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_923, x = var_1074)[name = tensor("out_27")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([12, 1, 256])]; + tensor out_29 = reshape(shape = var_1078, x = out_27)[name = tensor("out_29")]; + tensor var_1080 = silu(x = input_169)[name = tensor("op_1080")]; + tensor input_171 = mul(x = var_1080, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_921, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1090 = const()[name = tensor("op_1090"), val = tensor([1, 12, 1, 256])]; + tensor var_1091 = reshape(shape = var_1090, x = xt_1)[name = tensor("op_1091")]; + tensor var_1092_perm_0 = const()[name = tensor("op_1092_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1095 = const()[name = tensor("op_1095"), val = tensor([1, 12, 256])]; + tensor var_1092 = transpose(perm = var_1092_perm_0, x = var_1091)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1095, x = var_1092)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1118 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 1, 3, 256])]; + tensor var_1120 = reshape(shape = concat_1, x = var_1118)[name = tensor("op_1120")]; + tensor var_1121_axes_0 = const()[name = tensor("op_1121_axes_0"), val = tensor([0])]; + tensor var_1121 = expand_dims(axes = var_1121_axes_0, x = var_1120)[name = tensor("op_1121")]; + tensor var_1122_perm_0 = const()[name = tensor("op_1122_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1123_axes_0 = const()[name = tensor("op_1123_axes_0"), val = tensor([-2])]; + tensor var_1122 = transpose(perm = var_1122_perm_0, x = var_1121)[name = tensor("transpose_21")]; + tensor var_1123 = squeeze(axes = var_1123_axes_0, x = var_1122)[name = tensor("op_1123")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 1, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1123)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 1, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1123)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 1, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1123)[name = tensor("v_11")]; + tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([12, 4, 64])]; + tensor var_1132 = reshape(shape = var_1131, x = q_11)[name = tensor("op_1132")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1138 = const()[name = tensor("op_1138"), val = tensor([12, 4, 64])]; + tensor var_1139 = reshape(shape = var_1138, x = k_11)[name = tensor("op_1139")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([12, 4, 64])]; + tensor var_1146 = reshape(shape = var_1145, x = v_11)[name = tensor("op_1146")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1149 = const()[name = tensor("op_1149"), val = tensor([1, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1132)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1149, x = q_13)[name = tensor("q_15")]; + tensor var_1151 = const()[name = tensor("op_1151"), val = tensor([1, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1139)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1151, x = k_13)[name = tensor("k_15")]; + tensor var_1153 = const()[name = tensor("op_1153"), val = tensor([1, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1146)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1153, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([2, 0, 1, 3])]; + tensor var_1161 = const()[name = tensor("op_1161"), val = tensor([12, 256])]; + tensor var_1157 = transpose(perm = var_1156, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1161, x = var_1157)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1165 = const()[name = tensor("op_1165"), val = tensor([12, 1, 256])]; + tensor attn_output_7 = reshape(shape = var_1165, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_921, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_921, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 12, 256])]; + tensor x_31 = reshape(shape = var_1185, x = xt_3)[name = tensor("x_31")]; + tensor var_1187_perm_0 = const()[name = tensor("op_1187_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([12, 1, 256])]; + tensor var_1187 = transpose(perm = var_1187_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1191, x = var_1187)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1199 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1200 = const()[name = tensor("op_1200"), val = tensor([12, 1, 4, 64])]; + tensor var_1201 = reshape(shape = var_1200, x = var_1199)[name = tensor("op_1201")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1205 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1206 = const()[name = tensor("op_1206"), val = tensor(0x1p-3)]; + tensor var_1207 = mul(x = var_1205, y = var_1206)[name = tensor("op_1207")]; + tensor var_1208 = const()[name = tensor("op_1208"), val = tensor([12, 1, 4, 64])]; + tensor var_1209 = reshape(shape = var_1208, x = var_1207)[name = tensor("op_1209")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1213 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1214 = const()[name = tensor("op_1214"), val = tensor([12, 1, 4, 64])]; + tensor var_1215 = reshape(shape = var_1214, x = var_1213)[name = tensor("op_1215")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_926, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1209)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1201)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 1])]; + tensor var_1231 = reshape(shape = var_1230, x = sqrt_s_t)[name = tensor("op_1231")]; + tensor M = real_div(x = var_1047, y = var_1231)[name = tensor("M")]; + tensor var_1233 = mul(x = qk, y = M)[name = tensor("op_1233")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1215)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1233, y = v_17)[name = tensor("inner")]; + tensor var_1235_transpose_x_0 = const()[name = tensor("op_1235_transpose_x_0"), val = tensor(false)]; + tensor var_1235_transpose_y_0 = const()[name = tensor("op_1235_transpose_y_0"), val = tensor(false)]; + tensor var_1235 = matmul(transpose_x = var_1235_transpose_x_0, transpose_y = var_1235_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1235")]; + tensor var_1236 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1236")]; + tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1, 1, 1])]; + tensor var_1238 = reshape(shape = var_1237, x = var_1236)[name = tensor("op_1238")]; + tensor cross = mul(x = var_1235, y = var_1238)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1061)[name = tensor("v_masked")]; + tensor var_1244 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1244")]; + tensor var_1246_transpose_x_1 = const()[name = tensor("op_1246_transpose_x_1"), val = tensor(true)]; + tensor var_1246_transpose_y_1 = const()[name = tensor("op_1246_transpose_y_1"), val = tensor(false)]; + tensor var_1246 = matmul(transpose_x = var_1246_transpose_x_1, transpose_y = var_1246_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1246")]; + tensor new_kv_unnorm = add(x = var_1244, y = var_1246)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1069)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_926, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1255_perm_0 = const()[name = tensor("op_1255_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1255 = transpose(perm = var_1255_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_923, x = var_1255)[name = tensor("out_33")]; + tensor var_1259 = const()[name = tensor("op_1259"), val = tensor([12, 1, 256])]; + tensor out = reshape(shape = var_1259, x = out_33)[name = tensor("out")]; + tensor var_1261 = silu(x = input_187)[name = tensor("op_1261")]; + tensor input_189 = mul(x = var_1261, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_921, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1271 = const()[name = tensor("op_1271"), val = tensor([1, 12, 1, 256])]; + tensor var_1272 = reshape(shape = var_1271, x = xt_5)[name = tensor("op_1272")]; + tensor var_1273_perm_0 = const()[name = tensor("op_1273_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1276 = const()[name = tensor("op_1276"), val = tensor([1, 12, 256])]; + tensor var_1273 = transpose(perm = var_1273_perm_0, x = var_1272)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1276, x = var_1273)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1299 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 1, 3, 256])]; + tensor var_1301 = reshape(shape = concat_2, x = var_1299)[name = tensor("op_1301")]; + tensor var_1302_axes_0 = const()[name = tensor("op_1302_axes_0"), val = tensor([0])]; + tensor var_1302 = expand_dims(axes = var_1302_axes_0, x = var_1301)[name = tensor("op_1302")]; + tensor var_1303_perm_0 = const()[name = tensor("op_1303_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1304_axes_0 = const()[name = tensor("op_1304_axes_0"), val = tensor([-2])]; + tensor var_1303 = transpose(perm = var_1303_perm_0, x = var_1302)[name = tensor("transpose_8")]; + tensor var_1304 = squeeze(axes = var_1304_axes_0, x = var_1303)[name = tensor("op_1304")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 1, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1304)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 1, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1304)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 1, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1304)[name = tensor("v_19")]; + tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([12, 4, 64])]; + tensor var_1313 = reshape(shape = var_1312, x = q_19)[name = tensor("op_1313")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1319 = const()[name = tensor("op_1319"), val = tensor([12, 4, 64])]; + tensor var_1320 = reshape(shape = var_1319, x = k_19)[name = tensor("op_1320")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([12, 4, 64])]; + tensor var_1327 = reshape(shape = var_1326, x = v_19)[name = tensor("op_1327")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1330 = const()[name = tensor("op_1330"), val = tensor([1, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1313)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1330, x = q_21)[name = tensor("q")]; + tensor var_1332 = const()[name = tensor("op_1332"), val = tensor([1, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1320)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1332, x = k_21)[name = tensor("k")]; + tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1327)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1334, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([2, 0, 1, 3])]; + tensor var_1342 = const()[name = tensor("op_1342"), val = tensor([12, 256])]; + tensor var_1338 = transpose(perm = var_1337, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1342, x = var_1338)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1346 = const()[name = tensor("op_1346"), val = tensor([12, 1, 256])]; + tensor attn_output = reshape(shape = var_1346, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_921, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_921, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 12, 256])]; + tensor input = reshape(shape = var_1366, x = xt)[name = tensor("input")]; + tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([-1])]; + tensor var_1369 = reduce_l2_norm(axes = var_1368, keep_dims = var_924, x = input)[name = tensor("op_1369")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_916, beta = const_42, x = var_1369)[name = tensor("clip_5")]; + tensor var_1371 = real_div(x = input, y = clip_5)[name = tensor("op_1371")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([1, 256, 12])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1371)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = emb, y = reshape_1)[name = tensor("matmul_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 1, 11])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = matmul_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1375")]; + tensor var_1377_axis_0 = const()[name = tensor("op_1377_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1377_axis_0, values = (var_1073, nkv))[name = tensor("op_1377")]; + tensor var_1379_axis_0 = const()[name = tensor("op_1379_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1379_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1379")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/weights/weight.bin b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c296c51a2b2c4945fe134faa770e07d217e44f25 --- /dev/null +++ b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4387c7c31bad0c5b3ca91db8c92f0e25fa701bf2cf5ba182f8edc39a0a51825b +size 44407680 diff --git a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..2b188342f30379c189c2ef1ec6d955b9033ba14d --- /dev/null +++ b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0e1bf629771d4cad61a66e7932eb77dae2f0231edb3f0070409798b2d0c643c +size 171364 diff --git a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c296c51a2b2c4945fe134faa770e07d217e44f25 --- /dev/null +++ b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4387c7c31bad0c5b3ca91db8c92f0e25fa701bf2cf5ba182f8edc39a0a51825b +size 44407680 diff --git a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Manifest.json b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..84f6f3de60aa59bdc9ffe7ddbf987d76c5f22beb --- /dev/null +++ b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "377FDFC9-ABF8-4C96-8917-8CBDA39D46CD": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "EEC7467E-63A3-4C31-9B54-74606F035FAA": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "377FDFC9-ABF8-4C96-8917-8CBDA39D46CD" +} diff --git a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/analytics/coremldata.bin b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..397635a9a0e1cd59d8cf6642abba267486b68f69 --- /dev/null +++ b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1b0bab18a5f1a234fd3b8f86af03830354b0ff7d715753b069b582fd0dac6de +size 243 diff --git a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/coremldata.bin b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..e804d1c5d44837686767f82aef6a0244dc12e26c --- /dev/null +++ b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76832edb1250c28bb6886d9fd633957adb70b61d6525ec6e5ee1131c50305350 +size 1308 diff --git a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/metadata.json b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..fcec287e71dc3d3cfdc7a74070099ad04b2b2ae4 --- /dev/null +++ b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND DIHARD II streaming diarizer (pipeline, T=2, max_speakers=10)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 2 × 10)", + "shortDescription" : "", + "shape" : "[1, 2, 10]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 68, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 48, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 14, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 47 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 2 × 345)", + "shortDescription" : "", + "shape" : "[1, 2, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2)", + "shortDescription" : "", + "shape" : "[2]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"dih2\", \"model_label\": \"DIHARD II\", \"variant\": \"pipeline\", \"chunk_size\": 2, \"step_duration_ms\": 200, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_dih2_200ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/model.mil b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..5651eaf71d9adaeffd8f5bfe34de4f9950314977 --- /dev/null +++ b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/model.mil @@ -0,0 +1,1246 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 2, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 2, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 2, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([2, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 2, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 2, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; + tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; + tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; + tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; + tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; + tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 1, 256])]; + tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, true, true])]; + tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; + tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; + tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; + tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; + tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5))[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_257_split_sizes_0 = const()[name = tensor("op_257_split_sizes_0"), val = tensor([256, 256])]; + tensor var_257_axis_0 = const()[name = tensor("op_257_axis_0"), val = tensor(1)]; + tensor var_257_0, tensor var_257_1 = split(axis = var_257_axis_0, split_sizes = var_257_split_sizes_0, x = inputs_3)[name = tensor("op_257")]; + tensor var_259 = sigmoid(x = var_257_1)[name = tensor("op_259")]; + tensor inputs_5 = mul(x = var_257_0, y = var_259)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([2, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_290_begin_0 = const()[name = tensor("op_290_begin_0"), val = tensor([0, -1, 0])]; + tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([2, 16, 256])]; + tensor var_290_end_mask_0 = const()[name = tensor("op_290_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_290 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, x = conv_out_1)[name = tensor("op_290")]; + tensor var_292_perm_0 = const()[name = tensor("op_292_perm_0"), val = tensor([1, 0, 2])]; + tensor var_292 = transpose(perm = var_292_perm_0, x = var_290)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_292)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_315 = const()[name = tensor("op_315"), val = tensor(0x1p-1)]; + tensor var_316 = mul(x = input_39, y = var_315)[name = tensor("op_316")]; + tensor input_41 = add(x = var_316, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_345 = const()[name = tensor("op_345"), val = tensor(0x1p-1)]; + tensor var_346 = mul(x = input_51, y = var_345)[name = tensor("op_346")]; + tensor input_53 = add(x = var_346, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_360 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 2, 4, 64])]; + tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_366 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_367 = const()[name = tensor("op_367"), val = tensor(0x1p-3)]; + tensor var_368 = mul(x = var_366, y = var_367)[name = tensor("op_368")]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 2, 4, 64])]; + tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_374 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_375 = const()[name = tensor("op_375"), val = tensor([1, 2, 4, 64])]; + tensor var_376 = reshape(shape = var_375, x = var_374)[name = tensor("op_376")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_370)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_362)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([2, 1])]; + tensor var_387 = reshape(shape = var_386, x = sqrt_s_t_3)[name = tensor("op_387")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_387)[name = tensor("M_3")]; + tensor var_389 = mul(x = qk_3, y = M_3)[name = tensor("op_389")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_376)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_389, y = v_3)[name = tensor("inner_3")]; + tensor var_391_transpose_x_0 = const()[name = tensor("op_391_transpose_x_0"), val = tensor(false)]; + tensor var_391_transpose_y_0 = const()[name = tensor("op_391_transpose_y_0"), val = tensor(false)]; + tensor var_391 = matmul(transpose_x = var_391_transpose_x_0, transpose_y = var_391_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_391")]; + tensor var_392 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_392")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor([1, 1, 2, 1])]; + tensor var_394 = reshape(shape = var_393, x = var_392)[name = tensor("op_394")]; + tensor cross_3 = mul(x = var_391, y = var_394)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_397 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_397")]; + tensor var_399_transpose_x_1 = const()[name = tensor("op_399_transpose_x_1"), val = tensor(true)]; + tensor var_399_transpose_y_1 = const()[name = tensor("op_399_transpose_y_1"), val = tensor(false)]; + tensor var_399 = matmul(transpose_x = var_399_transpose_x_1, transpose_y = var_399_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_399")]; + tensor new_kv_unnorm_3 = add(x = var_397, y = var_399)[name = tensor("new_kv_unnorm_3")]; + tensor var_401 = const()[name = tensor("op_401"), val = tensor(0x1p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_401)[name = tensor("new_scale_3")]; + tensor var_403 = sqrt(x = new_scale_3)[name = tensor("op_403")]; + tensor var_404 = real_div(x = new_kv_unnorm_3, y = var_403)[name = tensor("op_404")]; + tensor var_405_perm_0 = const()[name = tensor("op_405_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_405 = transpose(perm = var_405_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_405)[name = tensor("out_9")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 2, 256])]; + tensor out_11 = reshape(shape = var_409, x = out_9)[name = tensor("out_11")]; + tensor var_411 = silu(x = input_57)[name = tensor("op_411")]; + tensor input_59 = mul(x = var_411, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_7_begin_0 = const()[name = tensor("window_7_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_7_end_0 = const()[name = tensor("window_7_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_7_end_mask_0 = const()[name = tensor("window_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_7_squeeze_mask_0 = const()[name = tensor("window_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_7 = slice_by_index(begin = window_7_begin_0, end = window_7_end_0, end_mask = window_7_end_mask_0, squeeze_mask = window_7_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_7")]; + tensor var_419_begin_0 = const()[name = tensor("op_419_begin_0"), val = tensor([0, 0, 0])]; + tensor var_419_end_0 = const()[name = tensor("op_419_end_0"), val = tensor([1, 1, 256])]; + tensor var_419_end_mask_0 = const()[name = tensor("op_419_end_mask_0"), val = tensor([true, false, true])]; + tensor var_419 = slice_by_index(begin = var_419_begin_0, end = var_419_end_0, end_mask = var_419_end_mask_0, x = x_9)[name = tensor("op_419")]; + tensor var_422_begin_0 = const()[name = tensor("op_422_begin_0"), val = tensor([0, 1, 0])]; + tensor var_422_end_0 = const()[name = tensor("op_422_end_0"), val = tensor([1, 16, 256])]; + tensor var_422_end_mask_0 = const()[name = tensor("op_422_end_mask_0"), val = tensor([true, true, true])]; + tensor var_422 = slice_by_index(begin = var_422_begin_0, end = var_422_end_0, end_mask = var_422_end_mask_0, x = window_7)[name = tensor("op_422")]; + tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; + tensor window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_422, var_419))[name = tensor("window_9")]; + tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 1, 0])]; + tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; + tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, true, true])]; + tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; + tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; + tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; + tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; + tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_9, window_11))[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_455_split_sizes_0 = const()[name = tensor("op_455_split_sizes_0"), val = tensor([256, 256])]; + tensor var_455_axis_0 = const()[name = tensor("op_455_axis_0"), val = tensor(1)]; + tensor var_455_0, tensor var_455_1 = split(axis = var_455_axis_0, split_sizes = var_455_split_sizes_0, x = inputs_13)[name = tensor("op_455")]; + tensor var_457 = sigmoid(x = var_455_1)[name = tensor("op_457")]; + tensor inputs_15 = mul(x = var_455_0, y = var_457)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([2, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_488_begin_0 = const()[name = tensor("op_488_begin_0"), val = tensor([0, -1, 0])]; + tensor var_488_end_0 = const()[name = tensor("op_488_end_0"), val = tensor([2, 16, 256])]; + tensor var_488_end_mask_0 = const()[name = tensor("op_488_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_488 = slice_by_index(begin = var_488_begin_0, end = var_488_end_0, end_mask = var_488_end_mask_0, x = conv_out_3)[name = tensor("op_488")]; + tensor var_490_perm_0 = const()[name = tensor("op_490_perm_0"), val = tensor([1, 0, 2])]; + tensor var_490 = transpose(perm = var_490_perm_0, x = var_488)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_490)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_513 = const()[name = tensor("op_513"), val = tensor(0x1p-1)]; + tensor var_514 = mul(x = input_79, y = var_513)[name = tensor("op_514")]; + tensor input_81 = add(x = var_514, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_543 = const()[name = tensor("op_543"), val = tensor(0x1p-1)]; + tensor var_544 = mul(x = input_91, y = var_543)[name = tensor("op_544")]; + tensor input_93 = add(x = var_544, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_558 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_559 = const()[name = tensor("op_559"), val = tensor([1, 2, 4, 64])]; + tensor var_560 = reshape(shape = var_559, x = var_558)[name = tensor("op_560")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_564 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_565 = const()[name = tensor("op_565"), val = tensor(0x1p-3)]; + tensor var_566 = mul(x = var_564, y = var_565)[name = tensor("op_566")]; + tensor var_567 = const()[name = tensor("op_567"), val = tensor([1, 2, 4, 64])]; + tensor var_568 = reshape(shape = var_567, x = var_566)[name = tensor("op_568")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_572 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 2, 4, 64])]; + tensor var_574 = reshape(shape = var_573, x = var_572)[name = tensor("op_574")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_568)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_560)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_584 = const()[name = tensor("op_584"), val = tensor([2, 1])]; + tensor var_585 = reshape(shape = var_584, x = sqrt_s_t_5)[name = tensor("op_585")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_585)[name = tensor("M_5")]; + tensor var_587 = mul(x = qk_5, y = M_5)[name = tensor("op_587")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_574)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_587, y = v_5)[name = tensor("inner_5")]; + tensor var_589_transpose_x_0 = const()[name = tensor("op_589_transpose_x_0"), val = tensor(false)]; + tensor var_589_transpose_y_0 = const()[name = tensor("op_589_transpose_y_0"), val = tensor(false)]; + tensor var_589 = matmul(transpose_x = var_589_transpose_x_0, transpose_y = var_589_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_589")]; + tensor var_590 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_590")]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 2, 1])]; + tensor var_592 = reshape(shape = var_591, x = var_590)[name = tensor("op_592")]; + tensor cross_5 = mul(x = var_589, y = var_592)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_595 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_595")]; + tensor var_597_transpose_x_1 = const()[name = tensor("op_597_transpose_x_1"), val = tensor(true)]; + tensor var_597_transpose_y_1 = const()[name = tensor("op_597_transpose_y_1"), val = tensor(false)]; + tensor var_597 = matmul(transpose_x = var_597_transpose_x_1, transpose_y = var_597_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_597")]; + tensor new_kv_unnorm_5 = add(x = var_595, y = var_597)[name = tensor("new_kv_unnorm_5")]; + tensor var_599 = const()[name = tensor("op_599"), val = tensor(0x1p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_599)[name = tensor("new_scale_5")]; + tensor var_601 = sqrt(x = new_scale_5)[name = tensor("op_601")]; + tensor var_602 = real_div(x = new_kv_unnorm_5, y = var_601)[name = tensor("op_602")]; + tensor var_603_perm_0 = const()[name = tensor("op_603_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_603 = transpose(perm = var_603_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_603)[name = tensor("out_15")]; + tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 2, 256])]; + tensor out_17 = reshape(shape = var_607, x = out_15)[name = tensor("out_17")]; + tensor var_609 = silu(x = input_97)[name = tensor("op_609")]; + tensor input_99 = mul(x = var_609, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; + tensor var_617_begin_0 = const()[name = tensor("op_617_begin_0"), val = tensor([0, 0, 0])]; + tensor var_617_end_0 = const()[name = tensor("op_617_end_0"), val = tensor([1, 1, 256])]; + tensor var_617_end_mask_0 = const()[name = tensor("op_617_end_mask_0"), val = tensor([true, false, true])]; + tensor var_617 = slice_by_index(begin = var_617_begin_0, end = var_617_end_0, end_mask = var_617_end_mask_0, x = x_15)[name = tensor("op_617")]; + tensor var_620_begin_0 = const()[name = tensor("op_620_begin_0"), val = tensor([0, 1, 0])]; + tensor var_620_end_0 = const()[name = tensor("op_620_end_0"), val = tensor([1, 16, 256])]; + tensor var_620_end_mask_0 = const()[name = tensor("op_620_end_mask_0"), val = tensor([true, true, true])]; + tensor var_620 = slice_by_index(begin = var_620_begin_0, end = var_620_end_0, end_mask = var_620_end_mask_0, x = window_13)[name = tensor("op_620")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_620, var_617))[name = tensor("window_15")]; + tensor var_625_begin_0 = const()[name = tensor("op_625_begin_0"), val = tensor([0, 1, 0])]; + tensor var_625_end_0 = const()[name = tensor("op_625_end_0"), val = tensor([1, 1, 256])]; + tensor var_625_end_mask_0 = const()[name = tensor("op_625_end_mask_0"), val = tensor([true, true, true])]; + tensor var_625 = slice_by_index(begin = var_625_begin_0, end = var_625_end_0, end_mask = var_625_end_mask_0, x = x_15)[name = tensor("op_625")]; + tensor var_628_begin_0 = const()[name = tensor("op_628_begin_0"), val = tensor([0, 1, 0])]; + tensor var_628_end_0 = const()[name = tensor("op_628_end_0"), val = tensor([1, 16, 256])]; + tensor var_628_end_mask_0 = const()[name = tensor("op_628_end_mask_0"), val = tensor([true, true, true])]; + tensor var_628 = slice_by_index(begin = var_628_begin_0, end = var_628_end_0, end_mask = var_628_end_mask_0, x = window_15)[name = tensor("op_628")]; + tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; + tensor window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_628, var_625))[name = tensor("window_17")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_15, window_17))[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_653_split_sizes_0 = const()[name = tensor("op_653_split_sizes_0"), val = tensor([256, 256])]; + tensor var_653_axis_0 = const()[name = tensor("op_653_axis_0"), val = tensor(1)]; + tensor var_653_0, tensor var_653_1 = split(axis = var_653_axis_0, split_sizes = var_653_split_sizes_0, x = inputs_23)[name = tensor("op_653")]; + tensor var_655 = sigmoid(x = var_653_1)[name = tensor("op_655")]; + tensor inputs_25 = mul(x = var_653_0, y = var_655)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([2, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_686_begin_0 = const()[name = tensor("op_686_begin_0"), val = tensor([0, -1, 0])]; + tensor var_686_end_0 = const()[name = tensor("op_686_end_0"), val = tensor([2, 16, 256])]; + tensor var_686_end_mask_0 = const()[name = tensor("op_686_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_686 = slice_by_index(begin = var_686_begin_0, end = var_686_end_0, end_mask = var_686_end_mask_0, x = conv_out_5)[name = tensor("op_686")]; + tensor var_688_perm_0 = const()[name = tensor("op_688_perm_0"), val = tensor([1, 0, 2])]; + tensor var_688 = transpose(perm = var_688_perm_0, x = var_686)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_688)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_711 = const()[name = tensor("op_711"), val = tensor(0x1p-1)]; + tensor var_712 = mul(x = input_119, y = var_711)[name = tensor("op_712")]; + tensor input_121 = add(x = var_712, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_741 = const()[name = tensor("op_741"), val = tensor(0x1p-1)]; + tensor var_742 = mul(x = input_131, y = var_741)[name = tensor("op_742")]; + tensor input_133 = add(x = var_742, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_756 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_757 = const()[name = tensor("op_757"), val = tensor([1, 2, 4, 64])]; + tensor var_758 = reshape(shape = var_757, x = var_756)[name = tensor("op_758")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_762 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_763 = const()[name = tensor("op_763"), val = tensor(0x1p-3)]; + tensor var_764 = mul(x = var_762, y = var_763)[name = tensor("op_764")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 2, 4, 64])]; + tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_770 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 2, 4, 64])]; + tensor var_772 = reshape(shape = var_771, x = var_770)[name = tensor("op_772")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_766)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_758)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_782 = const()[name = tensor("op_782"), val = tensor([2, 1])]; + tensor var_783 = reshape(shape = var_782, x = sqrt_s_t_7)[name = tensor("op_783")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_783)[name = tensor("M_7")]; + tensor var_785 = mul(x = qk_7, y = M_7)[name = tensor("op_785")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_772)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_785, y = v_7)[name = tensor("inner_7")]; + tensor var_787_transpose_x_0 = const()[name = tensor("op_787_transpose_x_0"), val = tensor(false)]; + tensor var_787_transpose_y_0 = const()[name = tensor("op_787_transpose_y_0"), val = tensor(false)]; + tensor var_787 = matmul(transpose_x = var_787_transpose_x_0, transpose_y = var_787_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_787")]; + tensor var_788 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_788")]; + tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 1, 2, 1])]; + tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; + tensor cross_7 = mul(x = var_787, y = var_790)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_793 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_793")]; + tensor var_795_transpose_x_1 = const()[name = tensor("op_795_transpose_x_1"), val = tensor(true)]; + tensor var_795_transpose_y_1 = const()[name = tensor("op_795_transpose_y_1"), val = tensor(false)]; + tensor var_795 = matmul(transpose_x = var_795_transpose_x_1, transpose_y = var_795_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_795")]; + tensor new_kv_unnorm_7 = add(x = var_793, y = var_795)[name = tensor("new_kv_unnorm_7")]; + tensor var_797 = const()[name = tensor("op_797"), val = tensor(0x1p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_797)[name = tensor("new_scale_7")]; + tensor var_799 = sqrt(x = new_scale_7)[name = tensor("op_799")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_799)[name = tensor("nkv_1")]; + tensor var_801_perm_0 = const()[name = tensor("op_801_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_801 = transpose(perm = var_801_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_801)[name = tensor("out_21")]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 2, 256])]; + tensor out_23 = reshape(shape = var_805, x = out_21)[name = tensor("out_23")]; + tensor var_807 = silu(x = input_137)[name = tensor("op_807")]; + tensor input_139 = mul(x = var_807, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_19_begin_0 = const()[name = tensor("window_19_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_19_end_0 = const()[name = tensor("window_19_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_19_end_mask_0 = const()[name = tensor("window_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_19_squeeze_mask_0 = const()[name = tensor("window_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_19 = slice_by_index(begin = window_19_begin_0, end = window_19_end_0, end_mask = window_19_end_mask_0, squeeze_mask = window_19_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_19")]; + tensor var_815_begin_0 = const()[name = tensor("op_815_begin_0"), val = tensor([0, 0, 0])]; + tensor var_815_end_0 = const()[name = tensor("op_815_end_0"), val = tensor([1, 1, 256])]; + tensor var_815_end_mask_0 = const()[name = tensor("op_815_end_mask_0"), val = tensor([true, false, true])]; + tensor var_815 = slice_by_index(begin = var_815_begin_0, end = var_815_end_0, end_mask = var_815_end_mask_0, x = x_21)[name = tensor("op_815")]; + tensor var_818_begin_0 = const()[name = tensor("op_818_begin_0"), val = tensor([0, 1, 0])]; + tensor var_818_end_0 = const()[name = tensor("op_818_end_0"), val = tensor([1, 16, 256])]; + tensor var_818_end_mask_0 = const()[name = tensor("op_818_end_mask_0"), val = tensor([true, true, true])]; + tensor var_818 = slice_by_index(begin = var_818_begin_0, end = var_818_end_0, end_mask = var_818_end_mask_0, x = window_19)[name = tensor("op_818")]; + tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; + tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_818, var_815))[name = tensor("window_21")]; + tensor var_823_begin_0 = const()[name = tensor("op_823_begin_0"), val = tensor([0, 1, 0])]; + tensor var_823_end_0 = const()[name = tensor("op_823_end_0"), val = tensor([1, 1, 256])]; + tensor var_823_end_mask_0 = const()[name = tensor("op_823_end_mask_0"), val = tensor([true, true, true])]; + tensor var_823 = slice_by_index(begin = var_823_begin_0, end = var_823_end_0, end_mask = var_823_end_mask_0, x = x_21)[name = tensor("op_823")]; + tensor var_826_begin_0 = const()[name = tensor("op_826_begin_0"), val = tensor([0, 1, 0])]; + tensor var_826_end_0 = const()[name = tensor("op_826_end_0"), val = tensor([1, 16, 256])]; + tensor var_826_end_mask_0 = const()[name = tensor("op_826_end_mask_0"), val = tensor([true, true, true])]; + tensor var_826 = slice_by_index(begin = var_826_begin_0, end = var_826_end_0, end_mask = var_826_end_mask_0, x = window_21)[name = tensor("op_826")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_826, var_823))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_21, window))[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_851_split_sizes_0 = const()[name = tensor("op_851_split_sizes_0"), val = tensor([256, 256])]; + tensor var_851_axis_0 = const()[name = tensor("op_851_axis_0"), val = tensor(1)]; + tensor var_851_0, tensor var_851_1 = split(axis = var_851_axis_0, split_sizes = var_851_split_sizes_0, x = inputs_33)[name = tensor("op_851")]; + tensor var_853 = sigmoid(x = var_851_1)[name = tensor("op_853")]; + tensor inputs_35 = mul(x = var_851_0, y = var_853)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([2, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_884_begin_0 = const()[name = tensor("op_884_begin_0"), val = tensor([0, -1, 0])]; + tensor var_884_end_0 = const()[name = tensor("op_884_end_0"), val = tensor([2, 16, 256])]; + tensor var_884_end_mask_0 = const()[name = tensor("op_884_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_884 = slice_by_index(begin = var_884_begin_0, end = var_884_end_0, end_mask = var_884_end_mask_0, x = conv_out_7)[name = tensor("op_884")]; + tensor var_886_perm_0 = const()[name = tensor("op_886_perm_0"), val = tensor([1, 0, 2])]; + tensor var_886 = transpose(perm = var_886_perm_0, x = var_884)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_886)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_909 = const()[name = tensor("op_909"), val = tensor(0x1p-1)]; + tensor var_910 = mul(x = input_159, y = var_909)[name = tensor("op_910")]; + tensor input_161 = add(x = var_910, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_928_begin_0 = const()[name = tensor("op_928_begin_0"), val = tensor([0, 0, 2])]; + tensor var_928_end_0 = const()[name = tensor("op_928_end_0"), val = tensor([1, 256, 20])]; + tensor var_928_end_mask_0 = const()[name = tensor("op_928_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_928_begin_0, end = var_928_end_0, end_mask = var_928_end_mask_0, x = cat)[name = tensor("op_928")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_930 = const()[name = tensor("op_930"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_931 = reduce_l2_norm(axes = var_930, keep_dims = var_29, x = input_163)[name = tensor("op_931")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_931)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_935_axis_0 = const()[name = tensor("op_935_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_935_axis_0, values = (var_206, var_404, var_602, nkv_1))[name = tensor("op_935")]; + tensor var_937_axis_0 = const()[name = tensor("op_937_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_937_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_937")]; + tensor var_939_axis_0 = const()[name = tensor("op_939_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_939_axis_0, values = (window_5, window_11, window_17, window))[name = tensor("op_939")]; + tensor var_948 = const()[name = tensor("op_948"), val = tensor(0x1.5798eep-27)]; + tensor var_953 = const()[name = tensor("op_953"), val = tensor(0x1.4f8b58p-17)]; + tensor var_955 = const()[name = tensor("op_955"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_956 = const()[name = tensor("op_956"), val = tensor(true)]; + tensor var_958 = const()[name = tensor("op_958"), val = tensor(0x1p+0)]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor(-1)]; + tensor var_968 = const()[name = tensor("op_968"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_1030_axes_0 = const()[name = tensor("op_1030_axes_0"), val = tensor([2])]; + tensor var_1030 = expand_dims(axes = var_1030_axes_0, x = emb)[name = tensor("op_1030")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1030)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_962, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1038_perm_0 = const()[name = tensor("op_1038_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1042 = const()[name = tensor("op_1042"), val = tensor([12, 2, 256])]; + tensor var_1038 = transpose(perm = var_1038_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1042, x = var_1038)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1050 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1051 = const()[name = tensor("op_1051"), val = tensor([12, 2, 4, 64])]; + tensor var_1052 = reshape(shape = var_1051, x = var_1050)[name = tensor("op_1052")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1056 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1057 = const()[name = tensor("op_1057"), val = tensor(0x1p-3)]; + tensor var_1058 = mul(x = var_1056, y = var_1057)[name = tensor("op_1058")]; + tensor var_1059 = const()[name = tensor("op_1059"), val = tensor([12, 2, 4, 64])]; + tensor var_1060 = reshape(shape = var_1059, x = var_1058)[name = tensor("op_1060")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1064 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1065 = const()[name = tensor("op_1065"), val = tensor([12, 2, 4, 64])]; + tensor var_1066 = reshape(shape = var_1065, x = var_1064)[name = tensor("op_1066")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_968, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_958, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1060)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1052)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([1, 2])]; + tensor var_1079 = reshape(shape = var_1078, x = valid_mask)[name = tensor("op_1079")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1079)[name = tensor("causal_with_valid_1")]; + tensor var_1081 = const()[name = tensor("op_1081"), val = tensor([2, 1])]; + tensor var_1082 = reshape(shape = var_1081, x = sqrt_s_t_9)[name = tensor("op_1082")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1082)[name = tensor("M_9")]; + tensor var_1084 = mul(x = qk_9, y = M_9)[name = tensor("op_1084")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1066)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1084, y = v_9)[name = tensor("inner_9")]; + tensor var_1086_transpose_x_0 = const()[name = tensor("op_1086_transpose_x_0"), val = tensor(false)]; + tensor var_1086_transpose_y_0 = const()[name = tensor("op_1086_transpose_y_0"), val = tensor(false)]; + tensor var_1086 = matmul(transpose_x = var_1086_transpose_x_0, transpose_y = var_1086_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1086")]; + tensor var_1087 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1087")]; + tensor var_1088 = const()[name = tensor("op_1088"), val = tensor([1, 1, 2, 1])]; + tensor var_1089 = reshape(shape = var_1088, x = var_1087)[name = tensor("op_1089")]; + tensor cross_9 = mul(x = var_1086, y = var_1089)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([1, 1, 2, 1])]; + tensor var_1093 = reshape(shape = var_1092, x = valid_mask)[name = tensor("op_1093")]; + tensor v_masked_1 = mul(x = v_9, y = var_1093)[name = tensor("v_masked_1")]; + tensor var_1095 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1095")]; + tensor var_1097_transpose_x_1 = const()[name = tensor("op_1097_transpose_x_1"), val = tensor(true)]; + tensor var_1097_transpose_y_1 = const()[name = tensor("op_1097_transpose_y_1"), val = tensor(false)]; + tensor var_1097 = matmul(transpose_x = var_1097_transpose_x_1, transpose_y = var_1097_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1097")]; + tensor new_kv_unnorm_9 = add(x = var_1095, y = var_1097)[name = tensor("new_kv_unnorm_9")]; + tensor var_1099_keep_dims_0 = const()[name = tensor("op_1099_keep_dims_0"), val = tensor(false)]; + tensor var_1099 = reduce_sum(keep_dims = var_1099_keep_dims_0, x = valid_mask)[name = tensor("op_1099")]; + tensor var_1100 = const()[name = tensor("op_1100"), val = tensor([1])]; + tensor var_1101 = reshape(shape = var_1100, x = var_1099)[name = tensor("op_1101")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1101)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_958, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1105 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1105")]; + tensor var_1106_perm_0 = const()[name = tensor("op_1106_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1106 = transpose(perm = var_1106_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_955, x = var_1106)[name = tensor("out_27")]; + tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([12, 2, 256])]; + tensor out_29 = reshape(shape = var_1110, x = out_27)[name = tensor("out_29")]; + tensor var_1112 = silu(x = input_169)[name = tensor("op_1112")]; + tensor input_171 = mul(x = var_1112, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_953, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1122 = const()[name = tensor("op_1122"), val = tensor([1, 12, 2, 256])]; + tensor var_1123 = reshape(shape = var_1122, x = xt_1)[name = tensor("op_1123")]; + tensor var_1124_perm_0 = const()[name = tensor("op_1124_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([2, 12, 256])]; + tensor var_1124 = transpose(perm = var_1124_perm_0, x = var_1123)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1127, x = var_1124)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1150 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 2, 3, 256])]; + tensor var_1152 = reshape(shape = concat_1, x = var_1150)[name = tensor("op_1152")]; + tensor var_1153_axes_0 = const()[name = tensor("op_1153_axes_0"), val = tensor([0])]; + tensor var_1153 = expand_dims(axes = var_1153_axes_0, x = var_1152)[name = tensor("op_1153")]; + tensor var_1154_perm_0 = const()[name = tensor("op_1154_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1155_axes_0 = const()[name = tensor("op_1155_axes_0"), val = tensor([-2])]; + tensor var_1154 = transpose(perm = var_1154_perm_0, x = var_1153)[name = tensor("transpose_21")]; + tensor var_1155 = squeeze(axes = var_1155_axes_0, x = var_1154)[name = tensor("op_1155")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 2, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1155)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 2, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1155)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 2, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1155)[name = tensor("v_11")]; + tensor var_1163 = const()[name = tensor("op_1163"), val = tensor([12, 8, 64])]; + tensor var_1164 = reshape(shape = var_1163, x = q_11)[name = tensor("op_1164")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1170 = const()[name = tensor("op_1170"), val = tensor([12, 8, 64])]; + tensor var_1171 = reshape(shape = var_1170, x = k_11)[name = tensor("op_1171")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor([12, 8, 64])]; + tensor var_1178 = reshape(shape = var_1177, x = v_11)[name = tensor("op_1178")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1181 = const()[name = tensor("op_1181"), val = tensor([2, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1164)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1181, x = q_13)[name = tensor("q_15")]; + tensor var_1183 = const()[name = tensor("op_1183"), val = tensor([2, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1171)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1183, x = k_13)[name = tensor("k_15")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([2, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1178)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1185, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1188 = const()[name = tensor("op_1188"), val = tensor([2, 0, 1, 3])]; + tensor var_1193 = const()[name = tensor("op_1193"), val = tensor([24, 256])]; + tensor var_1189 = transpose(perm = var_1188, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1193, x = var_1189)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([12, 2, 256])]; + tensor attn_output_7 = reshape(shape = var_1197, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_953, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_953, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([1, 2, 12, 256])]; + tensor x_31 = reshape(shape = var_1217, x = xt_3)[name = tensor("x_31")]; + tensor var_1219_perm_0 = const()[name = tensor("op_1219_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1223 = const()[name = tensor("op_1223"), val = tensor([12, 2, 256])]; + tensor var_1219 = transpose(perm = var_1219_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1223, x = var_1219)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1231 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([12, 2, 4, 64])]; + tensor var_1233 = reshape(shape = var_1232, x = var_1231)[name = tensor("op_1233")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1237 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor(0x1p-3)]; + tensor var_1239 = mul(x = var_1237, y = var_1238)[name = tensor("op_1239")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([12, 2, 4, 64])]; + tensor var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor("op_1241")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1245 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1246 = const()[name = tensor("op_1246"), val = tensor([12, 2, 4, 64])]; + tensor var_1247 = reshape(shape = var_1246, x = var_1245)[name = tensor("op_1247")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_958, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1241)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1233)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([2, 1])]; + tensor var_1263 = reshape(shape = var_1262, x = sqrt_s_t)[name = tensor("op_1263")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1263)[name = tensor("M")]; + tensor var_1265 = mul(x = qk, y = M)[name = tensor("op_1265")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1247)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1265, y = v_17)[name = tensor("inner")]; + tensor var_1267_transpose_x_0 = const()[name = tensor("op_1267_transpose_x_0"), val = tensor(false)]; + tensor var_1267_transpose_y_0 = const()[name = tensor("op_1267_transpose_y_0"), val = tensor(false)]; + tensor var_1267 = matmul(transpose_x = var_1267_transpose_x_0, transpose_y = var_1267_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1267")]; + tensor var_1268 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1268")]; + tensor var_1269 = const()[name = tensor("op_1269"), val = tensor([1, 1, 2, 1])]; + tensor var_1270 = reshape(shape = var_1269, x = var_1268)[name = tensor("op_1270")]; + tensor cross = mul(x = var_1267, y = var_1270)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1093)[name = tensor("v_masked")]; + tensor var_1276 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1276")]; + tensor var_1278_transpose_x_1 = const()[name = tensor("op_1278_transpose_x_1"), val = tensor(true)]; + tensor var_1278_transpose_y_1 = const()[name = tensor("op_1278_transpose_y_1"), val = tensor(false)]; + tensor var_1278 = matmul(transpose_x = var_1278_transpose_x_1, transpose_y = var_1278_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1278")]; + tensor new_kv_unnorm = add(x = var_1276, y = var_1278)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1101)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_958, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1287_perm_0 = const()[name = tensor("op_1287_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1287 = transpose(perm = var_1287_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_955, x = var_1287)[name = tensor("out_33")]; + tensor var_1291 = const()[name = tensor("op_1291"), val = tensor([12, 2, 256])]; + tensor out = reshape(shape = var_1291, x = out_33)[name = tensor("out")]; + tensor var_1293 = silu(x = input_187)[name = tensor("op_1293")]; + tensor input_189 = mul(x = var_1293, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_953, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1303 = const()[name = tensor("op_1303"), val = tensor([1, 12, 2, 256])]; + tensor var_1304 = reshape(shape = var_1303, x = xt_5)[name = tensor("op_1304")]; + tensor var_1305_perm_0 = const()[name = tensor("op_1305_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1308 = const()[name = tensor("op_1308"), val = tensor([2, 12, 256])]; + tensor var_1305 = transpose(perm = var_1305_perm_0, x = var_1304)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1308, x = var_1305)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1331 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 2, 3, 256])]; + tensor var_1333 = reshape(shape = concat_2, x = var_1331)[name = tensor("op_1333")]; + tensor var_1334_axes_0 = const()[name = tensor("op_1334_axes_0"), val = tensor([0])]; + tensor var_1334 = expand_dims(axes = var_1334_axes_0, x = var_1333)[name = tensor("op_1334")]; + tensor var_1335_perm_0 = const()[name = tensor("op_1335_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1336_axes_0 = const()[name = tensor("op_1336_axes_0"), val = tensor([-2])]; + tensor var_1335 = transpose(perm = var_1335_perm_0, x = var_1334)[name = tensor("transpose_8")]; + tensor var_1336 = squeeze(axes = var_1336_axes_0, x = var_1335)[name = tensor("op_1336")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 2, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1336)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 2, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1336)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 2, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1336)[name = tensor("v_19")]; + tensor var_1344 = const()[name = tensor("op_1344"), val = tensor([12, 8, 64])]; + tensor var_1345 = reshape(shape = var_1344, x = q_19)[name = tensor("op_1345")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1351 = const()[name = tensor("op_1351"), val = tensor([12, 8, 64])]; + tensor var_1352 = reshape(shape = var_1351, x = k_19)[name = tensor("op_1352")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([12, 8, 64])]; + tensor var_1359 = reshape(shape = var_1358, x = v_19)[name = tensor("op_1359")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1362 = const()[name = tensor("op_1362"), val = tensor([2, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1345)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1362, x = q_21)[name = tensor("q")]; + tensor var_1364 = const()[name = tensor("op_1364"), val = tensor([2, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1352)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1364, x = k_21)[name = tensor("k")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([2, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1359)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1366, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([2, 0, 1, 3])]; + tensor var_1374 = const()[name = tensor("op_1374"), val = tensor([24, 256])]; + tensor var_1370 = transpose(perm = var_1369, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1374, x = var_1370)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1378 = const()[name = tensor("op_1378"), val = tensor([12, 2, 256])]; + tensor attn_output = reshape(shape = var_1378, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_953, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_953, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([1, 2, 12, 256])]; + tensor input = reshape(shape = var_1398, x = xt)[name = tensor("input")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([-1])]; + tensor var_1401 = reduce_l2_norm(axes = var_1400, keep_dims = var_956, x = input)[name = tensor("op_1401")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_948, beta = const_42, x = var_1401)[name = tensor("clip_5")]; + tensor var_1403 = real_div(x = input, y = clip_5)[name = tensor("op_1403")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([2, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([2, 256, 12])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1403)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 2, 12])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 2, 11])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1407")]; + tensor var_1409_axis_0 = const()[name = tensor("op_1409_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1409_axis_0, values = (var_1105, nkv))[name = tensor("op_1409")]; + tensor var_1411_axis_0 = const()[name = tensor("op_1411_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1411_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1411")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/weights/weight.bin b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..b811c17b5f6597fe3c34ea4afc001e21310e354d --- /dev/null +++ b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52539be9f8e94ca3ae6e1b78b204acd878ad2c94c42ad35340e6e6c3aa259caf +size 44419968 diff --git a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..51b7851ed05bebc0051a092a84d8dcde58dd2ed9 --- /dev/null +++ b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2de52b68412da2b73e916665ebf41cce5eacd2010c99e37de2722c6ce40094a2 +size 179874 diff --git a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..b811c17b5f6597fe3c34ea4afc001e21310e354d --- /dev/null +++ b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52539be9f8e94ca3ae6e1b78b204acd878ad2c94c42ad35340e6e6c3aa259caf +size 44419968 diff --git a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Manifest.json b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..439619156a11723a4eb1eb25e302786f8f2998b0 --- /dev/null +++ b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "07858CB5-65AA-49FE-837C-C3CED8A0B958": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "E331E090-E6BB-40EA-B66D-83ED2654BAA6": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "E331E090-E6BB-40EA-B66D-83ED2654BAA6" +} diff --git a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/analytics/coremldata.bin b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..aa2356ddff010ed9351b885712064ec8c4872b59 --- /dev/null +++ b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0e2c05f0fa28ecf43578092711f1d39b3176b08681e17319b4ff36eac9caff0 +size 243 diff --git a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/coremldata.bin b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..4bac83cc94d2161567ef0eaff6090eac0493e52a --- /dev/null +++ b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7894d3847f177d05bc7f718737302118c6e2735e776cb8d02cf573f134d75b4 +size 1308 diff --git a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/metadata.json b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..61a23767da180facd27d20100346d45d4522e1f9 --- /dev/null +++ b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND DIHARD II streaming diarizer (pipeline, T=3, max_speakers=10)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 3 × 10)", + "shortDescription" : "", + "shape" : "[1, 3, 10]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 68, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 56, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 18, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 47 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 3 × 345)", + "shortDescription" : "", + "shape" : "[1, 3, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 3)", + "shortDescription" : "", + "shape" : "[3]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"dih2\", \"model_label\": \"DIHARD II\", \"variant\": \"pipeline\", \"chunk_size\": 3, \"step_duration_ms\": 300, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_dih2_300ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/model.mil b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..48f1c415ceeabfb14a7edfb748becb4366d6de9f --- /dev/null +++ b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/model.mil @@ -0,0 +1,1286 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 3, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 3, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 3, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([3, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 3, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1.8p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 3, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; + tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; + tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; + tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; + tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; + tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; + tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; + tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; + tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; + tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; + tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; + tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; + tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; + tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 1, 256])]; + tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, true, true])]; + tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; + tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; + tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; + tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; + tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7))[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_265_split_sizes_0 = const()[name = tensor("op_265_split_sizes_0"), val = tensor([256, 256])]; + tensor var_265_axis_0 = const()[name = tensor("op_265_axis_0"), val = tensor(1)]; + tensor var_265_0, tensor var_265_1 = split(axis = var_265_axis_0, split_sizes = var_265_split_sizes_0, x = inputs_3)[name = tensor("op_265")]; + tensor var_267 = sigmoid(x = var_265_1)[name = tensor("op_267")]; + tensor inputs_5 = mul(x = var_265_0, y = var_267)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([3, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_298_begin_0 = const()[name = tensor("op_298_begin_0"), val = tensor([0, -1, 0])]; + tensor var_298_end_0 = const()[name = tensor("op_298_end_0"), val = tensor([3, 16, 256])]; + tensor var_298_end_mask_0 = const()[name = tensor("op_298_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_298 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, x = conv_out_1)[name = tensor("op_298")]; + tensor var_300_perm_0 = const()[name = tensor("op_300_perm_0"), val = tensor([1, 0, 2])]; + tensor var_300 = transpose(perm = var_300_perm_0, x = var_298)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_300)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_323 = const()[name = tensor("op_323"), val = tensor(0x1p-1)]; + tensor var_324 = mul(x = input_39, y = var_323)[name = tensor("op_324")]; + tensor input_41 = add(x = var_324, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_353 = const()[name = tensor("op_353"), val = tensor(0x1p-1)]; + tensor var_354 = mul(x = input_51, y = var_353)[name = tensor("op_354")]; + tensor input_53 = add(x = var_354, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_368 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 3, 4, 64])]; + tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_374 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_375 = const()[name = tensor("op_375"), val = tensor(0x1p-3)]; + tensor var_376 = mul(x = var_374, y = var_375)[name = tensor("op_376")]; + tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 3, 4, 64])]; + tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_382 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_383 = const()[name = tensor("op_383"), val = tensor([1, 3, 4, 64])]; + tensor var_384 = reshape(shape = var_383, x = var_382)[name = tensor("op_384")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_378)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_370)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_394 = const()[name = tensor("op_394"), val = tensor([3, 1])]; + tensor var_395 = reshape(shape = var_394, x = sqrt_s_t_3)[name = tensor("op_395")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_395)[name = tensor("M_3")]; + tensor var_397 = mul(x = qk_3, y = M_3)[name = tensor("op_397")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_384)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_397, y = v_3)[name = tensor("inner_3")]; + tensor var_399_transpose_x_0 = const()[name = tensor("op_399_transpose_x_0"), val = tensor(false)]; + tensor var_399_transpose_y_0 = const()[name = tensor("op_399_transpose_y_0"), val = tensor(false)]; + tensor var_399 = matmul(transpose_x = var_399_transpose_x_0, transpose_y = var_399_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_399")]; + tensor var_400 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_400")]; + tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 3, 1])]; + tensor var_402 = reshape(shape = var_401, x = var_400)[name = tensor("op_402")]; + tensor cross_3 = mul(x = var_399, y = var_402)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_405 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_405")]; + tensor var_407_transpose_x_1 = const()[name = tensor("op_407_transpose_x_1"), val = tensor(true)]; + tensor var_407_transpose_y_1 = const()[name = tensor("op_407_transpose_y_1"), val = tensor(false)]; + tensor var_407 = matmul(transpose_x = var_407_transpose_x_1, transpose_y = var_407_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_407")]; + tensor new_kv_unnorm_3 = add(x = var_405, y = var_407)[name = tensor("new_kv_unnorm_3")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor(0x1.8p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_409)[name = tensor("new_scale_3")]; + tensor var_411 = sqrt(x = new_scale_3)[name = tensor("op_411")]; + tensor var_412 = real_div(x = new_kv_unnorm_3, y = var_411)[name = tensor("op_412")]; + tensor var_413_perm_0 = const()[name = tensor("op_413_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_413 = transpose(perm = var_413_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_413)[name = tensor("out_9")]; + tensor var_417 = const()[name = tensor("op_417"), val = tensor([1, 3, 256])]; + tensor out_11 = reshape(shape = var_417, x = out_9)[name = tensor("out_11")]; + tensor var_419 = silu(x = input_57)[name = tensor("op_419")]; + tensor input_59 = mul(x = var_419, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; + tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 0, 0])]; + tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; + tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, false, true])]; + tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; + tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; + tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; + tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; + tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; + tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 1, 0])]; + tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 2, 256])]; + tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; + tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; + tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; + tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; + tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; + tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; + tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; + tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 2, 0])]; + tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 1, 256])]; + tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, true, true])]; + tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; + tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; + tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; + tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; + tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_11, window_13, window_15))[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_471_split_sizes_0 = const()[name = tensor("op_471_split_sizes_0"), val = tensor([256, 256])]; + tensor var_471_axis_0 = const()[name = tensor("op_471_axis_0"), val = tensor(1)]; + tensor var_471_0, tensor var_471_1 = split(axis = var_471_axis_0, split_sizes = var_471_split_sizes_0, x = inputs_13)[name = tensor("op_471")]; + tensor var_473 = sigmoid(x = var_471_1)[name = tensor("op_473")]; + tensor inputs_15 = mul(x = var_471_0, y = var_473)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([3, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_504_begin_0 = const()[name = tensor("op_504_begin_0"), val = tensor([0, -1, 0])]; + tensor var_504_end_0 = const()[name = tensor("op_504_end_0"), val = tensor([3, 16, 256])]; + tensor var_504_end_mask_0 = const()[name = tensor("op_504_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_504 = slice_by_index(begin = var_504_begin_0, end = var_504_end_0, end_mask = var_504_end_mask_0, x = conv_out_3)[name = tensor("op_504")]; + tensor var_506_perm_0 = const()[name = tensor("op_506_perm_0"), val = tensor([1, 0, 2])]; + tensor var_506 = transpose(perm = var_506_perm_0, x = var_504)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_506)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_529 = const()[name = tensor("op_529"), val = tensor(0x1p-1)]; + tensor var_530 = mul(x = input_79, y = var_529)[name = tensor("op_530")]; + tensor input_81 = add(x = var_530, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_559 = const()[name = tensor("op_559"), val = tensor(0x1p-1)]; + tensor var_560 = mul(x = input_91, y = var_559)[name = tensor("op_560")]; + tensor input_93 = add(x = var_560, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_574 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 3, 4, 64])]; + tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_580 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_581 = const()[name = tensor("op_581"), val = tensor(0x1p-3)]; + tensor var_582 = mul(x = var_580, y = var_581)[name = tensor("op_582")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 3, 4, 64])]; + tensor var_584 = reshape(shape = var_583, x = var_582)[name = tensor("op_584")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_588 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_589 = const()[name = tensor("op_589"), val = tensor([1, 3, 4, 64])]; + tensor var_590 = reshape(shape = var_589, x = var_588)[name = tensor("op_590")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_584)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_576)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_600 = const()[name = tensor("op_600"), val = tensor([3, 1])]; + tensor var_601 = reshape(shape = var_600, x = sqrt_s_t_5)[name = tensor("op_601")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_601)[name = tensor("M_5")]; + tensor var_603 = mul(x = qk_5, y = M_5)[name = tensor("op_603")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_590)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_603, y = v_5)[name = tensor("inner_5")]; + tensor var_605_transpose_x_0 = const()[name = tensor("op_605_transpose_x_0"), val = tensor(false)]; + tensor var_605_transpose_y_0 = const()[name = tensor("op_605_transpose_y_0"), val = tensor(false)]; + tensor var_605 = matmul(transpose_x = var_605_transpose_x_0, transpose_y = var_605_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_605")]; + tensor var_606 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_606")]; + tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 1, 3, 1])]; + tensor var_608 = reshape(shape = var_607, x = var_606)[name = tensor("op_608")]; + tensor cross_5 = mul(x = var_605, y = var_608)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_611 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_611")]; + tensor var_613_transpose_x_1 = const()[name = tensor("op_613_transpose_x_1"), val = tensor(true)]; + tensor var_613_transpose_y_1 = const()[name = tensor("op_613_transpose_y_1"), val = tensor(false)]; + tensor var_613 = matmul(transpose_x = var_613_transpose_x_1, transpose_y = var_613_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_613")]; + tensor new_kv_unnorm_5 = add(x = var_611, y = var_613)[name = tensor("new_kv_unnorm_5")]; + tensor var_615 = const()[name = tensor("op_615"), val = tensor(0x1.8p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_615)[name = tensor("new_scale_5")]; + tensor var_617 = sqrt(x = new_scale_5)[name = tensor("op_617")]; + tensor var_618 = real_div(x = new_kv_unnorm_5, y = var_617)[name = tensor("op_618")]; + tensor var_619_perm_0 = const()[name = tensor("op_619_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_619 = transpose(perm = var_619_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_619)[name = tensor("out_15")]; + tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 3, 256])]; + tensor out_17 = reshape(shape = var_623, x = out_15)[name = tensor("out_17")]; + tensor var_625 = silu(x = input_97)[name = tensor("op_625")]; + tensor input_99 = mul(x = var_625, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_17_begin_0 = const()[name = tensor("window_17_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_17_end_0 = const()[name = tensor("window_17_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_17_end_mask_0 = const()[name = tensor("window_17_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_17_squeeze_mask_0 = const()[name = tensor("window_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_17 = slice_by_index(begin = window_17_begin_0, end = window_17_end_0, end_mask = window_17_end_mask_0, squeeze_mask = window_17_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_17")]; + tensor var_633_begin_0 = const()[name = tensor("op_633_begin_0"), val = tensor([0, 0, 0])]; + tensor var_633_end_0 = const()[name = tensor("op_633_end_0"), val = tensor([1, 1, 256])]; + tensor var_633_end_mask_0 = const()[name = tensor("op_633_end_mask_0"), val = tensor([true, false, true])]; + tensor var_633 = slice_by_index(begin = var_633_begin_0, end = var_633_end_0, end_mask = var_633_end_mask_0, x = x_15)[name = tensor("op_633")]; + tensor var_636_begin_0 = const()[name = tensor("op_636_begin_0"), val = tensor([0, 1, 0])]; + tensor var_636_end_0 = const()[name = tensor("op_636_end_0"), val = tensor([1, 16, 256])]; + tensor var_636_end_mask_0 = const()[name = tensor("op_636_end_mask_0"), val = tensor([true, true, true])]; + tensor var_636 = slice_by_index(begin = var_636_begin_0, end = var_636_end_0, end_mask = var_636_end_mask_0, x = window_17)[name = tensor("op_636")]; + tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; + tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_636, var_633))[name = tensor("window_19")]; + tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 1, 0])]; + tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 2, 256])]; + tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, false, true])]; + tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = x_15)[name = tensor("op_641")]; + tensor var_644_begin_0 = const()[name = tensor("op_644_begin_0"), val = tensor([0, 1, 0])]; + tensor var_644_end_0 = const()[name = tensor("op_644_end_0"), val = tensor([1, 16, 256])]; + tensor var_644_end_mask_0 = const()[name = tensor("op_644_end_mask_0"), val = tensor([true, true, true])]; + tensor var_644 = slice_by_index(begin = var_644_begin_0, end = var_644_end_0, end_mask = var_644_end_mask_0, x = window_19)[name = tensor("op_644")]; + tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; + tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_644, var_641))[name = tensor("window_21")]; + tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 2, 0])]; + tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; + tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, true, true])]; + tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; + tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; + tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; + tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; + tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; + tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_19, window_21, window_23))[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_677_split_sizes_0 = const()[name = tensor("op_677_split_sizes_0"), val = tensor([256, 256])]; + tensor var_677_axis_0 = const()[name = tensor("op_677_axis_0"), val = tensor(1)]; + tensor var_677_0, tensor var_677_1 = split(axis = var_677_axis_0, split_sizes = var_677_split_sizes_0, x = inputs_23)[name = tensor("op_677")]; + tensor var_679 = sigmoid(x = var_677_1)[name = tensor("op_679")]; + tensor inputs_25 = mul(x = var_677_0, y = var_679)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([3, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_710_begin_0 = const()[name = tensor("op_710_begin_0"), val = tensor([0, -1, 0])]; + tensor var_710_end_0 = const()[name = tensor("op_710_end_0"), val = tensor([3, 16, 256])]; + tensor var_710_end_mask_0 = const()[name = tensor("op_710_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_710 = slice_by_index(begin = var_710_begin_0, end = var_710_end_0, end_mask = var_710_end_mask_0, x = conv_out_5)[name = tensor("op_710")]; + tensor var_712_perm_0 = const()[name = tensor("op_712_perm_0"), val = tensor([1, 0, 2])]; + tensor var_712 = transpose(perm = var_712_perm_0, x = var_710)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_712)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_735 = const()[name = tensor("op_735"), val = tensor(0x1p-1)]; + tensor var_736 = mul(x = input_119, y = var_735)[name = tensor("op_736")]; + tensor input_121 = add(x = var_736, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor(0x1p-1)]; + tensor var_766 = mul(x = input_131, y = var_765)[name = tensor("op_766")]; + tensor input_133 = add(x = var_766, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_780 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 3, 4, 64])]; + tensor var_782 = reshape(shape = var_781, x = var_780)[name = tensor("op_782")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_786 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_787 = const()[name = tensor("op_787"), val = tensor(0x1p-3)]; + tensor var_788 = mul(x = var_786, y = var_787)[name = tensor("op_788")]; + tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 3, 4, 64])]; + tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_794 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_795 = const()[name = tensor("op_795"), val = tensor([1, 3, 4, 64])]; + tensor var_796 = reshape(shape = var_795, x = var_794)[name = tensor("op_796")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_790)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_782)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_806 = const()[name = tensor("op_806"), val = tensor([3, 1])]; + tensor var_807 = reshape(shape = var_806, x = sqrt_s_t_7)[name = tensor("op_807")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_807)[name = tensor("M_7")]; + tensor var_809 = mul(x = qk_7, y = M_7)[name = tensor("op_809")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_796)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_809, y = v_7)[name = tensor("inner_7")]; + tensor var_811_transpose_x_0 = const()[name = tensor("op_811_transpose_x_0"), val = tensor(false)]; + tensor var_811_transpose_y_0 = const()[name = tensor("op_811_transpose_y_0"), val = tensor(false)]; + tensor var_811 = matmul(transpose_x = var_811_transpose_x_0, transpose_y = var_811_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_811")]; + tensor var_812 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_812")]; + tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 1, 3, 1])]; + tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; + tensor cross_7 = mul(x = var_811, y = var_814)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_817 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_817")]; + tensor var_819_transpose_x_1 = const()[name = tensor("op_819_transpose_x_1"), val = tensor(true)]; + tensor var_819_transpose_y_1 = const()[name = tensor("op_819_transpose_y_1"), val = tensor(false)]; + tensor var_819 = matmul(transpose_x = var_819_transpose_x_1, transpose_y = var_819_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_819")]; + tensor new_kv_unnorm_7 = add(x = var_817, y = var_819)[name = tensor("new_kv_unnorm_7")]; + tensor var_821 = const()[name = tensor("op_821"), val = tensor(0x1.8p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_821)[name = tensor("new_scale_7")]; + tensor var_823 = sqrt(x = new_scale_7)[name = tensor("op_823")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_823)[name = tensor("nkv_1")]; + tensor var_825_perm_0 = const()[name = tensor("op_825_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_825 = transpose(perm = var_825_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_825)[name = tensor("out_21")]; + tensor var_829 = const()[name = tensor("op_829"), val = tensor([1, 3, 256])]; + tensor out_23 = reshape(shape = var_829, x = out_21)[name = tensor("out_23")]; + tensor var_831 = silu(x = input_137)[name = tensor("op_831")]; + tensor input_139 = mul(x = var_831, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; + tensor var_839_begin_0 = const()[name = tensor("op_839_begin_0"), val = tensor([0, 0, 0])]; + tensor var_839_end_0 = const()[name = tensor("op_839_end_0"), val = tensor([1, 1, 256])]; + tensor var_839_end_mask_0 = const()[name = tensor("op_839_end_mask_0"), val = tensor([true, false, true])]; + tensor var_839 = slice_by_index(begin = var_839_begin_0, end = var_839_end_0, end_mask = var_839_end_mask_0, x = x_21)[name = tensor("op_839")]; + tensor var_842_begin_0 = const()[name = tensor("op_842_begin_0"), val = tensor([0, 1, 0])]; + tensor var_842_end_0 = const()[name = tensor("op_842_end_0"), val = tensor([1, 16, 256])]; + tensor var_842_end_mask_0 = const()[name = tensor("op_842_end_mask_0"), val = tensor([true, true, true])]; + tensor var_842 = slice_by_index(begin = var_842_begin_0, end = var_842_end_0, end_mask = var_842_end_mask_0, x = window_25)[name = tensor("op_842")]; + tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; + tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_842, var_839))[name = tensor("window_27")]; + tensor var_847_begin_0 = const()[name = tensor("op_847_begin_0"), val = tensor([0, 1, 0])]; + tensor var_847_end_0 = const()[name = tensor("op_847_end_0"), val = tensor([1, 2, 256])]; + tensor var_847_end_mask_0 = const()[name = tensor("op_847_end_mask_0"), val = tensor([true, false, true])]; + tensor var_847 = slice_by_index(begin = var_847_begin_0, end = var_847_end_0, end_mask = var_847_end_mask_0, x = x_21)[name = tensor("op_847")]; + tensor var_850_begin_0 = const()[name = tensor("op_850_begin_0"), val = tensor([0, 1, 0])]; + tensor var_850_end_0 = const()[name = tensor("op_850_end_0"), val = tensor([1, 16, 256])]; + tensor var_850_end_mask_0 = const()[name = tensor("op_850_end_mask_0"), val = tensor([true, true, true])]; + tensor var_850 = slice_by_index(begin = var_850_begin_0, end = var_850_end_0, end_mask = var_850_end_mask_0, x = window_27)[name = tensor("op_850")]; + tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; + tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_850, var_847))[name = tensor("window_29")]; + tensor var_855_begin_0 = const()[name = tensor("op_855_begin_0"), val = tensor([0, 2, 0])]; + tensor var_855_end_0 = const()[name = tensor("op_855_end_0"), val = tensor([1, 1, 256])]; + tensor var_855_end_mask_0 = const()[name = tensor("op_855_end_mask_0"), val = tensor([true, true, true])]; + tensor var_855 = slice_by_index(begin = var_855_begin_0, end = var_855_end_0, end_mask = var_855_end_mask_0, x = x_21)[name = tensor("op_855")]; + tensor var_858_begin_0 = const()[name = tensor("op_858_begin_0"), val = tensor([0, 1, 0])]; + tensor var_858_end_0 = const()[name = tensor("op_858_end_0"), val = tensor([1, 16, 256])]; + tensor var_858_end_mask_0 = const()[name = tensor("op_858_end_mask_0"), val = tensor([true, true, true])]; + tensor var_858 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = window_29)[name = tensor("op_858")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_858, var_855))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_27, window_29, window))[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_883_split_sizes_0 = const()[name = tensor("op_883_split_sizes_0"), val = tensor([256, 256])]; + tensor var_883_axis_0 = const()[name = tensor("op_883_axis_0"), val = tensor(1)]; + tensor var_883_0, tensor var_883_1 = split(axis = var_883_axis_0, split_sizes = var_883_split_sizes_0, x = inputs_33)[name = tensor("op_883")]; + tensor var_885 = sigmoid(x = var_883_1)[name = tensor("op_885")]; + tensor inputs_35 = mul(x = var_883_0, y = var_885)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([3, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_916_begin_0 = const()[name = tensor("op_916_begin_0"), val = tensor([0, -1, 0])]; + tensor var_916_end_0 = const()[name = tensor("op_916_end_0"), val = tensor([3, 16, 256])]; + tensor var_916_end_mask_0 = const()[name = tensor("op_916_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_916 = slice_by_index(begin = var_916_begin_0, end = var_916_end_0, end_mask = var_916_end_mask_0, x = conv_out_7)[name = tensor("op_916")]; + tensor var_918_perm_0 = const()[name = tensor("op_918_perm_0"), val = tensor([1, 0, 2])]; + tensor var_918 = transpose(perm = var_918_perm_0, x = var_916)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_918)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_941 = const()[name = tensor("op_941"), val = tensor(0x1p-1)]; + tensor var_942 = mul(x = input_159, y = var_941)[name = tensor("op_942")]; + tensor input_161 = add(x = var_942, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_960_begin_0 = const()[name = tensor("op_960_begin_0"), val = tensor([0, 0, 3])]; + tensor var_960_end_0 = const()[name = tensor("op_960_end_0"), val = tensor([1, 256, 21])]; + tensor var_960_end_mask_0 = const()[name = tensor("op_960_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_960_begin_0, end = var_960_end_0, end_mask = var_960_end_mask_0, x = cat)[name = tensor("op_960")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_963 = reduce_l2_norm(axes = var_962, keep_dims = var_29, x = input_163)[name = tensor("op_963")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_963)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_967_axis_0, values = (var_206, var_412, var_618, nkv_1))[name = tensor("op_967")]; + tensor var_969_axis_0 = const()[name = tensor("op_969_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_969_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_969")]; + tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_971_axis_0, values = (window_7, window_15, window_23, window))[name = tensor("op_971")]; + tensor var_980 = const()[name = tensor("op_980"), val = tensor(0x1.5798eep-27)]; + tensor var_985 = const()[name = tensor("op_985"), val = tensor(0x1.4f8b58p-17)]; + tensor var_987 = const()[name = tensor("op_987"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_988 = const()[name = tensor("op_988"), val = tensor(true)]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor(0x1p+0)]; + tensor var_994 = const()[name = tensor("op_994"), val = tensor(-1)]; + tensor var_1000 = const()[name = tensor("op_1000"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_1062_axes_0 = const()[name = tensor("op_1062_axes_0"), val = tensor([2])]; + tensor var_1062 = expand_dims(axes = var_1062_axes_0, x = emb)[name = tensor("op_1062")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1062)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_994, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1070_perm_0 = const()[name = tensor("op_1070_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1074 = const()[name = tensor("op_1074"), val = tensor([12, 3, 256])]; + tensor var_1070 = transpose(perm = var_1070_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1074, x = var_1070)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1082 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1083 = const()[name = tensor("op_1083"), val = tensor([12, 3, 4, 64])]; + tensor var_1084 = reshape(shape = var_1083, x = var_1082)[name = tensor("op_1084")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1088 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1089 = const()[name = tensor("op_1089"), val = tensor(0x1p-3)]; + tensor var_1090 = mul(x = var_1088, y = var_1089)[name = tensor("op_1090")]; + tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([12, 3, 4, 64])]; + tensor var_1092 = reshape(shape = var_1091, x = var_1090)[name = tensor("op_1092")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1096 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([12, 3, 4, 64])]; + tensor var_1098 = reshape(shape = var_1097, x = var_1096)[name = tensor("op_1098")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_1000, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_990, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1092)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1084)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([1, 3])]; + tensor var_1111 = reshape(shape = var_1110, x = valid_mask)[name = tensor("op_1111")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1111)[name = tensor("causal_with_valid_1")]; + tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([3, 1])]; + tensor var_1114 = reshape(shape = var_1113, x = sqrt_s_t_9)[name = tensor("op_1114")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1114)[name = tensor("M_9")]; + tensor var_1116 = mul(x = qk_9, y = M_9)[name = tensor("op_1116")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1098)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1116, y = v_9)[name = tensor("inner_9")]; + tensor var_1118_transpose_x_0 = const()[name = tensor("op_1118_transpose_x_0"), val = tensor(false)]; + tensor var_1118_transpose_y_0 = const()[name = tensor("op_1118_transpose_y_0"), val = tensor(false)]; + tensor var_1118 = matmul(transpose_x = var_1118_transpose_x_0, transpose_y = var_1118_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1118")]; + tensor var_1119 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1119")]; + tensor var_1120 = const()[name = tensor("op_1120"), val = tensor([1, 1, 3, 1])]; + tensor var_1121 = reshape(shape = var_1120, x = var_1119)[name = tensor("op_1121")]; + tensor cross_9 = mul(x = var_1118, y = var_1121)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1124 = const()[name = tensor("op_1124"), val = tensor([1, 1, 3, 1])]; + tensor var_1125 = reshape(shape = var_1124, x = valid_mask)[name = tensor("op_1125")]; + tensor v_masked_1 = mul(x = v_9, y = var_1125)[name = tensor("v_masked_1")]; + tensor var_1127 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1127")]; + tensor var_1129_transpose_x_1 = const()[name = tensor("op_1129_transpose_x_1"), val = tensor(true)]; + tensor var_1129_transpose_y_1 = const()[name = tensor("op_1129_transpose_y_1"), val = tensor(false)]; + tensor var_1129 = matmul(transpose_x = var_1129_transpose_x_1, transpose_y = var_1129_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1129")]; + tensor new_kv_unnorm_9 = add(x = var_1127, y = var_1129)[name = tensor("new_kv_unnorm_9")]; + tensor var_1131_keep_dims_0 = const()[name = tensor("op_1131_keep_dims_0"), val = tensor(false)]; + tensor var_1131 = reduce_sum(keep_dims = var_1131_keep_dims_0, x = valid_mask)[name = tensor("op_1131")]; + tensor var_1132 = const()[name = tensor("op_1132"), val = tensor([1])]; + tensor var_1133 = reshape(shape = var_1132, x = var_1131)[name = tensor("op_1133")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1133)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_990, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1137 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1137")]; + tensor var_1138_perm_0 = const()[name = tensor("op_1138_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1138 = transpose(perm = var_1138_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_987, x = var_1138)[name = tensor("out_27")]; + tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([12, 3, 256])]; + tensor out_29 = reshape(shape = var_1142, x = out_27)[name = tensor("out_29")]; + tensor var_1144 = silu(x = input_169)[name = tensor("op_1144")]; + tensor input_171 = mul(x = var_1144, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_985, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 12, 3, 256])]; + tensor var_1155 = reshape(shape = var_1154, x = xt_1)[name = tensor("op_1155")]; + tensor var_1156_perm_0 = const()[name = tensor("op_1156_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1159 = const()[name = tensor("op_1159"), val = tensor([3, 12, 256])]; + tensor var_1156 = transpose(perm = var_1156_perm_0, x = var_1155)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1159, x = var_1156)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1182 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 3, 3, 256])]; + tensor var_1184 = reshape(shape = concat_1, x = var_1182)[name = tensor("op_1184")]; + tensor var_1185_axes_0 = const()[name = tensor("op_1185_axes_0"), val = tensor([0])]; + tensor var_1185 = expand_dims(axes = var_1185_axes_0, x = var_1184)[name = tensor("op_1185")]; + tensor var_1186_perm_0 = const()[name = tensor("op_1186_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1187_axes_0 = const()[name = tensor("op_1187_axes_0"), val = tensor([-2])]; + tensor var_1186 = transpose(perm = var_1186_perm_0, x = var_1185)[name = tensor("transpose_21")]; + tensor var_1187 = squeeze(axes = var_1187_axes_0, x = var_1186)[name = tensor("op_1187")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 3, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1187)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 3, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1187)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 3, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1187)[name = tensor("v_11")]; + tensor var_1195 = const()[name = tensor("op_1195"), val = tensor([12, 12, 64])]; + tensor var_1196 = reshape(shape = var_1195, x = q_11)[name = tensor("op_1196")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1202 = const()[name = tensor("op_1202"), val = tensor([12, 12, 64])]; + tensor var_1203 = reshape(shape = var_1202, x = k_11)[name = tensor("op_1203")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1209 = const()[name = tensor("op_1209"), val = tensor([12, 12, 64])]; + tensor var_1210 = reshape(shape = var_1209, x = v_11)[name = tensor("op_1210")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1213 = const()[name = tensor("op_1213"), val = tensor([3, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1196)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1213, x = q_13)[name = tensor("q_15")]; + tensor var_1215 = const()[name = tensor("op_1215"), val = tensor([3, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1203)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1215, x = k_13)[name = tensor("k_15")]; + tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([3, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1210)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1217, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1220 = const()[name = tensor("op_1220"), val = tensor([2, 0, 1, 3])]; + tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([36, 256])]; + tensor var_1221 = transpose(perm = var_1220, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1225, x = var_1221)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1229 = const()[name = tensor("op_1229"), val = tensor([12, 3, 256])]; + tensor attn_output_7 = reshape(shape = var_1229, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_985, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_985, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([1, 3, 12, 256])]; + tensor x_31 = reshape(shape = var_1249, x = xt_3)[name = tensor("x_31")]; + tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1255 = const()[name = tensor("op_1255"), val = tensor([12, 3, 256])]; + tensor var_1251 = transpose(perm = var_1251_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1255, x = var_1251)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1263 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1264 = const()[name = tensor("op_1264"), val = tensor([12, 3, 4, 64])]; + tensor var_1265 = reshape(shape = var_1264, x = var_1263)[name = tensor("op_1265")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1269 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1270 = const()[name = tensor("op_1270"), val = tensor(0x1p-3)]; + tensor var_1271 = mul(x = var_1269, y = var_1270)[name = tensor("op_1271")]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([12, 3, 4, 64])]; + tensor var_1273 = reshape(shape = var_1272, x = var_1271)[name = tensor("op_1273")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1277 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([12, 3, 4, 64])]; + tensor var_1279 = reshape(shape = var_1278, x = var_1277)[name = tensor("op_1279")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_990, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1273)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1265)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([3, 1])]; + tensor var_1295 = reshape(shape = var_1294, x = sqrt_s_t)[name = tensor("op_1295")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1295)[name = tensor("M")]; + tensor var_1297 = mul(x = qk, y = M)[name = tensor("op_1297")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1279)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1297, y = v_17)[name = tensor("inner")]; + tensor var_1299_transpose_x_0 = const()[name = tensor("op_1299_transpose_x_0"), val = tensor(false)]; + tensor var_1299_transpose_y_0 = const()[name = tensor("op_1299_transpose_y_0"), val = tensor(false)]; + tensor var_1299 = matmul(transpose_x = var_1299_transpose_x_0, transpose_y = var_1299_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1299")]; + tensor var_1300 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1300")]; + tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([1, 1, 3, 1])]; + tensor var_1302 = reshape(shape = var_1301, x = var_1300)[name = tensor("op_1302")]; + tensor cross = mul(x = var_1299, y = var_1302)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1125)[name = tensor("v_masked")]; + tensor var_1308 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1308")]; + tensor var_1310_transpose_x_1 = const()[name = tensor("op_1310_transpose_x_1"), val = tensor(true)]; + tensor var_1310_transpose_y_1 = const()[name = tensor("op_1310_transpose_y_1"), val = tensor(false)]; + tensor var_1310 = matmul(transpose_x = var_1310_transpose_x_1, transpose_y = var_1310_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1310")]; + tensor new_kv_unnorm = add(x = var_1308, y = var_1310)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1133)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_990, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1319_perm_0 = const()[name = tensor("op_1319_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1319 = transpose(perm = var_1319_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_987, x = var_1319)[name = tensor("out_33")]; + tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([12, 3, 256])]; + tensor out = reshape(shape = var_1323, x = out_33)[name = tensor("out")]; + tensor var_1325 = silu(x = input_187)[name = tensor("op_1325")]; + tensor input_189 = mul(x = var_1325, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_985, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor([1, 12, 3, 256])]; + tensor var_1336 = reshape(shape = var_1335, x = xt_5)[name = tensor("op_1336")]; + tensor var_1337_perm_0 = const()[name = tensor("op_1337_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1340 = const()[name = tensor("op_1340"), val = tensor([3, 12, 256])]; + tensor var_1337 = transpose(perm = var_1337_perm_0, x = var_1336)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1340, x = var_1337)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1363 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 3, 3, 256])]; + tensor var_1365 = reshape(shape = concat_2, x = var_1363)[name = tensor("op_1365")]; + tensor var_1366_axes_0 = const()[name = tensor("op_1366_axes_0"), val = tensor([0])]; + tensor var_1366 = expand_dims(axes = var_1366_axes_0, x = var_1365)[name = tensor("op_1366")]; + tensor var_1367_perm_0 = const()[name = tensor("op_1367_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1368_axes_0 = const()[name = tensor("op_1368_axes_0"), val = tensor([-2])]; + tensor var_1367 = transpose(perm = var_1367_perm_0, x = var_1366)[name = tensor("transpose_8")]; + tensor var_1368 = squeeze(axes = var_1368_axes_0, x = var_1367)[name = tensor("op_1368")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 3, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1368)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 3, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1368)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 3, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1368)[name = tensor("v_19")]; + tensor var_1376 = const()[name = tensor("op_1376"), val = tensor([12, 12, 64])]; + tensor var_1377 = reshape(shape = var_1376, x = q_19)[name = tensor("op_1377")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([12, 12, 64])]; + tensor var_1384 = reshape(shape = var_1383, x = k_19)[name = tensor("op_1384")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([12, 12, 64])]; + tensor var_1391 = reshape(shape = var_1390, x = v_19)[name = tensor("op_1391")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1394 = const()[name = tensor("op_1394"), val = tensor([3, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1377)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1394, x = q_21)[name = tensor("q")]; + tensor var_1396 = const()[name = tensor("op_1396"), val = tensor([3, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1384)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1396, x = k_21)[name = tensor("k")]; + tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([3, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1391)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1398, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1401 = const()[name = tensor("op_1401"), val = tensor([2, 0, 1, 3])]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([36, 256])]; + tensor var_1402 = transpose(perm = var_1401, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1406, x = var_1402)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1410 = const()[name = tensor("op_1410"), val = tensor([12, 3, 256])]; + tensor attn_output = reshape(shape = var_1410, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_985, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_985, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([1, 3, 12, 256])]; + tensor input = reshape(shape = var_1430, x = xt)[name = tensor("input")]; + tensor var_1432 = const()[name = tensor("op_1432"), val = tensor([-1])]; + tensor var_1433 = reduce_l2_norm(axes = var_1432, keep_dims = var_988, x = input)[name = tensor("op_1433")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_980, beta = const_42, x = var_1433)[name = tensor("clip_5")]; + tensor var_1435 = real_div(x = input, y = clip_5)[name = tensor("op_1435")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([3, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([3, 256, 12])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1435)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 3, 12])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 3, 11])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1439")]; + tensor var_1441_axis_0 = const()[name = tensor("op_1441_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1441_axis_0, values = (var_1137, nkv))[name = tensor("op_1441")]; + tensor var_1443_axis_0 = const()[name = tensor("op_1443_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1443_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1443")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/weights/weight.bin b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..196be541cdb161bd00f794476816a6da3c0044a5 --- /dev/null +++ b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a86c456707c856f7be42c00d33fdbb6bfabb878cff792434fffb8d47657290c +size 44432256 diff --git a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..667d39182df060a6ae205f7f1936b532292aaa00 --- /dev/null +++ b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9912c0d401a73181999617e3ae8cf8781300b6effc6ddd777a0d2d293c9dc351 +size 185467 diff --git a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..196be541cdb161bd00f794476816a6da3c0044a5 --- /dev/null +++ b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a86c456707c856f7be42c00d33fdbb6bfabb878cff792434fffb8d47657290c +size 44432256 diff --git a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Manifest.json b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..94cff4c78a7ab58f03319cc612dc624390995764 --- /dev/null +++ b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "655B25EA-0632-4B1A-B464-F7DFCE1FB184": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "E19F6AA3-6B6D-4C5E-8868-BE6F38C9CFE1": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "E19F6AA3-6B6D-4C5E-8868-BE6F38C9CFE1" +} diff --git a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/analytics/coremldata.bin b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..60912e32f95f143258d8e51db95a3edea8749f6c --- /dev/null +++ b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4159dfa0fd403ab65353f56fd6e2f87e53062263e1e136c98c4fcc19b79ffab8 +size 243 diff --git a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/coremldata.bin b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..410afba3c04fb8e61c00656e4d9170768e2bbe4a --- /dev/null +++ b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4deedfd08d987353bb9e2e6585c389e10e71fd41bcbe7d5b3140ff4ff9bf985 +size 1308 diff --git a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/metadata.json b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..c179a7eaf3aa7525369b7d58fdb519fbdc635371 --- /dev/null +++ b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND DIHARD II streaming diarizer (pipeline, T=4, max_speakers=10)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 4 × 10)", + "shortDescription" : "", + "shape" : "[1, 4, 10]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 68, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 64, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 22, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 47 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 4 × 345)", + "shortDescription" : "", + "shape" : "[1, 4, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4)", + "shortDescription" : "", + "shape" : "[4]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"dih2\", \"model_label\": \"DIHARD II\", \"variant\": \"pipeline\", \"chunk_size\": 4, \"step_duration_ms\": 400, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_dih2_400ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/model.mil b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..7135f6328b831d1038e863b851b7563d75bba5c3 --- /dev/null +++ b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/model.mil @@ -0,0 +1,1326 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982080)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983168)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336512)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337600)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338688)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339776)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340864)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345024)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393664)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394752)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443392)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444480)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445568)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446656)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708864)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709952)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972160)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973248)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235456)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236544)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498752)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499840)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762048)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763136)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764224)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766336)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290688)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307136)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308224)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309312)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310400)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311488)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312576)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574784)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575872)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576960)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581120)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629760)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630848)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679488)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680576)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681664)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682752)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683840)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688000)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736640)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737728)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786368)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787456)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788544)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789632)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051840)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052928)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315136)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316224)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578432)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579520)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841728)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842816)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105024)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106112)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107200)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109312)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633664)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650112)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651200)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652288)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653376)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654464)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655552)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917760)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918848)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919936)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924096)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972736)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973824)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022464)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023552)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024640)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025728)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026816)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030976)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079616)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080704)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129344)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130432)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131520)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132608)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394816)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395904)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658112)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659200)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921408)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922496)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184704)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185792)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448000)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449088)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450176)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452288)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976640)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993088)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994176)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995264)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996352)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997440)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998528)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260736)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261824)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262912)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267072)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315712)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316800)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365440)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366528)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367616)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368704)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369792)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373952)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422592)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423680)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472320)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473408)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474496)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475584)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737792)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738880)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001088)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002176)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264384)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265472)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527680)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528768)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790976)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792064)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793152)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795264)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319616)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336064)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337152)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338240)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339328)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340416)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341504)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603712)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604800)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605888)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610048)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658688)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659776)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708416)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709504)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710592)))]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710720)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711808)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236160)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237248)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499456)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500544)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762752)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763840)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026048)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027136)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289344)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290432)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552640)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553728)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554816)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555904)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818112)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821248)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607744)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608832)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609920)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618176)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715392)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716480)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813696)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814784)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815872)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816960)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079168)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080256)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342464)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343552)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605760)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606848)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869056)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870144)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132352)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133440)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134528)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135616)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397824)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400960)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187456)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188544)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189632)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197888)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295104)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296192)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393408)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394496)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 4, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 4, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 4, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([4, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 4, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 4, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; + tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; + tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; + tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; + tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; + tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; + tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; + tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; + tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; + tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; + tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; + tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; + tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; + tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 3, 256])]; + tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, false, true])]; + tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; + tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; + tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; + tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; + tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; + tensor var_245_begin_0 = const()[name = tensor("op_245_begin_0"), val = tensor([0, 3, 0])]; + tensor var_245_end_0 = const()[name = tensor("op_245_end_0"), val = tensor([1, 1, 256])]; + tensor var_245_end_mask_0 = const()[name = tensor("op_245_end_mask_0"), val = tensor([true, true, true])]; + tensor var_245 = slice_by_index(begin = var_245_begin_0, end = var_245_end_0, end_mask = var_245_end_mask_0, x = x_3)[name = tensor("op_245")]; + tensor var_248_begin_0 = const()[name = tensor("op_248_begin_0"), val = tensor([0, 1, 0])]; + tensor var_248_end_0 = const()[name = tensor("op_248_end_0"), val = tensor([1, 16, 256])]; + tensor var_248_end_mask_0 = const()[name = tensor("op_248_end_mask_0"), val = tensor([true, true, true])]; + tensor var_248 = slice_by_index(begin = var_248_begin_0, end = var_248_end_0, end_mask = var_248_end_mask_0, x = window_7)[name = tensor("op_248")]; + tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; + tensor window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_248, var_245))[name = tensor("window_9")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9))[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_273_split_sizes_0 = const()[name = tensor("op_273_split_sizes_0"), val = tensor([256, 256])]; + tensor var_273_axis_0 = const()[name = tensor("op_273_axis_0"), val = tensor(1)]; + tensor var_273_0, tensor var_273_1 = split(axis = var_273_axis_0, split_sizes = var_273_split_sizes_0, x = inputs_3)[name = tensor("op_273")]; + tensor var_275 = sigmoid(x = var_273_1)[name = tensor("op_275")]; + tensor inputs_5 = mul(x = var_273_0, y = var_275)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([4, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([0, -1, 0])]; + tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([4, 16, 256])]; + tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_306 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, x = conv_out_1)[name = tensor("op_306")]; + tensor var_308_perm_0 = const()[name = tensor("op_308_perm_0"), val = tensor([1, 0, 2])]; + tensor var_308 = transpose(perm = var_308_perm_0, x = var_306)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_308)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_331 = const()[name = tensor("op_331"), val = tensor(0x1p-1)]; + tensor var_332 = mul(x = input_39, y = var_331)[name = tensor("op_332")]; + tensor input_41 = add(x = var_332, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor(0x1p-1)]; + tensor var_362 = mul(x = input_51, y = var_361)[name = tensor("op_362")]; + tensor input_53 = add(x = var_362, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_376 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 4, 4, 64])]; + tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_382 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_383 = const()[name = tensor("op_383"), val = tensor(0x1p-3)]; + tensor var_384 = mul(x = var_382, y = var_383)[name = tensor("op_384")]; + tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 4, 4, 64])]; + tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_390 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 4, 4, 64])]; + tensor var_392 = reshape(shape = var_391, x = var_390)[name = tensor("op_392")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_386)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_378)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_402 = const()[name = tensor("op_402"), val = tensor([4, 1])]; + tensor var_403 = reshape(shape = var_402, x = sqrt_s_t_3)[name = tensor("op_403")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_403)[name = tensor("M_3")]; + tensor var_405 = mul(x = qk_3, y = M_3)[name = tensor("op_405")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_392)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_405, y = v_3)[name = tensor("inner_3")]; + tensor var_407_transpose_x_0 = const()[name = tensor("op_407_transpose_x_0"), val = tensor(false)]; + tensor var_407_transpose_y_0 = const()[name = tensor("op_407_transpose_y_0"), val = tensor(false)]; + tensor var_407 = matmul(transpose_x = var_407_transpose_x_0, transpose_y = var_407_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_407")]; + tensor var_408 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_408")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1, 4, 1])]; + tensor var_410 = reshape(shape = var_409, x = var_408)[name = tensor("op_410")]; + tensor cross_3 = mul(x = var_407, y = var_410)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_413 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_413")]; + tensor var_415_transpose_x_1 = const()[name = tensor("op_415_transpose_x_1"), val = tensor(true)]; + tensor var_415_transpose_y_1 = const()[name = tensor("op_415_transpose_y_1"), val = tensor(false)]; + tensor var_415 = matmul(transpose_x = var_415_transpose_x_1, transpose_y = var_415_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_415")]; + tensor new_kv_unnorm_3 = add(x = var_413, y = var_415)[name = tensor("new_kv_unnorm_3")]; + tensor var_417 = const()[name = tensor("op_417"), val = tensor(0x1p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_417)[name = tensor("new_scale_3")]; + tensor var_419 = sqrt(x = new_scale_3)[name = tensor("op_419")]; + tensor var_420 = real_div(x = new_kv_unnorm_3, y = var_419)[name = tensor("op_420")]; + tensor var_421_perm_0 = const()[name = tensor("op_421_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_421 = transpose(perm = var_421_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_421)[name = tensor("out_9")]; + tensor var_425 = const()[name = tensor("op_425"), val = tensor([1, 4, 256])]; + tensor out_11 = reshape(shape = var_425, x = out_9)[name = tensor("out_11")]; + tensor var_427 = silu(x = input_57)[name = tensor("op_427")]; + tensor input_59 = mul(x = var_427, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_11_begin_0 = const()[name = tensor("window_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_11_end_0 = const()[name = tensor("window_11_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_11_end_mask_0 = const()[name = tensor("window_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_11_squeeze_mask_0 = const()[name = tensor("window_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_11 = slice_by_index(begin = window_11_begin_0, end = window_11_end_0, end_mask = window_11_end_mask_0, squeeze_mask = window_11_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_11")]; + tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 0, 0])]; + tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 1, 256])]; + tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; + tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; + tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; + tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; + tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; + tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; + tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; + tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 1, 0])]; + tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 2, 256])]; + tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, false, true])]; + tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; + tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; + tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; + tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; + tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; + tensor var_451_begin_0 = const()[name = tensor("op_451_begin_0"), val = tensor([0, 2, 0])]; + tensor var_451_end_0 = const()[name = tensor("op_451_end_0"), val = tensor([1, 3, 256])]; + tensor var_451_end_mask_0 = const()[name = tensor("op_451_end_mask_0"), val = tensor([true, false, true])]; + tensor var_451 = slice_by_index(begin = var_451_begin_0, end = var_451_end_0, end_mask = var_451_end_mask_0, x = x_9)[name = tensor("op_451")]; + tensor var_454_begin_0 = const()[name = tensor("op_454_begin_0"), val = tensor([0, 1, 0])]; + tensor var_454_end_0 = const()[name = tensor("op_454_end_0"), val = tensor([1, 16, 256])]; + tensor var_454_end_mask_0 = const()[name = tensor("op_454_end_mask_0"), val = tensor([true, true, true])]; + tensor var_454 = slice_by_index(begin = var_454_begin_0, end = var_454_end_0, end_mask = var_454_end_mask_0, x = window_15)[name = tensor("op_454")]; + tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; + tensor window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_454, var_451))[name = tensor("window_17")]; + tensor var_459_begin_0 = const()[name = tensor("op_459_begin_0"), val = tensor([0, 3, 0])]; + tensor var_459_end_0 = const()[name = tensor("op_459_end_0"), val = tensor([1, 1, 256])]; + tensor var_459_end_mask_0 = const()[name = tensor("op_459_end_mask_0"), val = tensor([true, true, true])]; + tensor var_459 = slice_by_index(begin = var_459_begin_0, end = var_459_end_0, end_mask = var_459_end_mask_0, x = x_9)[name = tensor("op_459")]; + tensor var_462_begin_0 = const()[name = tensor("op_462_begin_0"), val = tensor([0, 1, 0])]; + tensor var_462_end_0 = const()[name = tensor("op_462_end_0"), val = tensor([1, 16, 256])]; + tensor var_462_end_mask_0 = const()[name = tensor("op_462_end_mask_0"), val = tensor([true, true, true])]; + tensor var_462 = slice_by_index(begin = var_462_begin_0, end = var_462_end_0, end_mask = var_462_end_mask_0, x = window_17)[name = tensor("op_462")]; + tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; + tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_462, var_459))[name = tensor("window_19")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_13, window_15, window_17, window_19))[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_487_split_sizes_0 = const()[name = tensor("op_487_split_sizes_0"), val = tensor([256, 256])]; + tensor var_487_axis_0 = const()[name = tensor("op_487_axis_0"), val = tensor(1)]; + tensor var_487_0, tensor var_487_1 = split(axis = var_487_axis_0, split_sizes = var_487_split_sizes_0, x = inputs_13)[name = tensor("op_487")]; + tensor var_489 = sigmoid(x = var_487_1)[name = tensor("op_489")]; + tensor inputs_15 = mul(x = var_487_0, y = var_489)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([4, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_520_begin_0 = const()[name = tensor("op_520_begin_0"), val = tensor([0, -1, 0])]; + tensor var_520_end_0 = const()[name = tensor("op_520_end_0"), val = tensor([4, 16, 256])]; + tensor var_520_end_mask_0 = const()[name = tensor("op_520_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_520 = slice_by_index(begin = var_520_begin_0, end = var_520_end_0, end_mask = var_520_end_mask_0, x = conv_out_3)[name = tensor("op_520")]; + tensor var_522_perm_0 = const()[name = tensor("op_522_perm_0"), val = tensor([1, 0, 2])]; + tensor var_522 = transpose(perm = var_522_perm_0, x = var_520)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_522)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_545 = const()[name = tensor("op_545"), val = tensor(0x1p-1)]; + tensor var_546 = mul(x = input_79, y = var_545)[name = tensor("op_546")]; + tensor input_81 = add(x = var_546, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-1)]; + tensor var_576 = mul(x = input_91, y = var_575)[name = tensor("op_576")]; + tensor input_93 = add(x = var_576, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_590 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 4, 4, 64])]; + tensor var_592 = reshape(shape = var_591, x = var_590)[name = tensor("op_592")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_596 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_597 = const()[name = tensor("op_597"), val = tensor(0x1p-3)]; + tensor var_598 = mul(x = var_596, y = var_597)[name = tensor("op_598")]; + tensor var_599 = const()[name = tensor("op_599"), val = tensor([1, 4, 4, 64])]; + tensor var_600 = reshape(shape = var_599, x = var_598)[name = tensor("op_600")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_604 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_605 = const()[name = tensor("op_605"), val = tensor([1, 4, 4, 64])]; + tensor var_606 = reshape(shape = var_605, x = var_604)[name = tensor("op_606")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_600)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_592)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_616 = const()[name = tensor("op_616"), val = tensor([4, 1])]; + tensor var_617 = reshape(shape = var_616, x = sqrt_s_t_5)[name = tensor("op_617")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_617)[name = tensor("M_5")]; + tensor var_619 = mul(x = qk_5, y = M_5)[name = tensor("op_619")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_606)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_619, y = v_5)[name = tensor("inner_5")]; + tensor var_621_transpose_x_0 = const()[name = tensor("op_621_transpose_x_0"), val = tensor(false)]; + tensor var_621_transpose_y_0 = const()[name = tensor("op_621_transpose_y_0"), val = tensor(false)]; + tensor var_621 = matmul(transpose_x = var_621_transpose_x_0, transpose_y = var_621_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_621")]; + tensor var_622 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_622")]; + tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 1, 4, 1])]; + tensor var_624 = reshape(shape = var_623, x = var_622)[name = tensor("op_624")]; + tensor cross_5 = mul(x = var_621, y = var_624)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_627 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_627")]; + tensor var_629_transpose_x_1 = const()[name = tensor("op_629_transpose_x_1"), val = tensor(true)]; + tensor var_629_transpose_y_1 = const()[name = tensor("op_629_transpose_y_1"), val = tensor(false)]; + tensor var_629 = matmul(transpose_x = var_629_transpose_x_1, transpose_y = var_629_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_629")]; + tensor new_kv_unnorm_5 = add(x = var_627, y = var_629)[name = tensor("new_kv_unnorm_5")]; + tensor var_631 = const()[name = tensor("op_631"), val = tensor(0x1p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_631)[name = tensor("new_scale_5")]; + tensor var_633 = sqrt(x = new_scale_5)[name = tensor("op_633")]; + tensor var_634 = real_div(x = new_kv_unnorm_5, y = var_633)[name = tensor("op_634")]; + tensor var_635_perm_0 = const()[name = tensor("op_635_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_635 = transpose(perm = var_635_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_635)[name = tensor("out_15")]; + tensor var_639 = const()[name = tensor("op_639"), val = tensor([1, 4, 256])]; + tensor out_17 = reshape(shape = var_639, x = out_15)[name = tensor("out_17")]; + tensor var_641 = silu(x = input_97)[name = tensor("op_641")]; + tensor input_99 = mul(x = var_641, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_21_begin_0 = const()[name = tensor("window_21_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_21_end_0 = const()[name = tensor("window_21_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_21_end_mask_0 = const()[name = tensor("window_21_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_21_squeeze_mask_0 = const()[name = tensor("window_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_21 = slice_by_index(begin = window_21_begin_0, end = window_21_end_0, end_mask = window_21_end_mask_0, squeeze_mask = window_21_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_21")]; + tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 0, 0])]; + tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; + tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, false, true])]; + tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; + tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; + tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; + tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; + tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; + tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; + tensor var_657_begin_0 = const()[name = tensor("op_657_begin_0"), val = tensor([0, 1, 0])]; + tensor var_657_end_0 = const()[name = tensor("op_657_end_0"), val = tensor([1, 2, 256])]; + tensor var_657_end_mask_0 = const()[name = tensor("op_657_end_mask_0"), val = tensor([true, false, true])]; + tensor var_657 = slice_by_index(begin = var_657_begin_0, end = var_657_end_0, end_mask = var_657_end_mask_0, x = x_15)[name = tensor("op_657")]; + tensor var_660_begin_0 = const()[name = tensor("op_660_begin_0"), val = tensor([0, 1, 0])]; + tensor var_660_end_0 = const()[name = tensor("op_660_end_0"), val = tensor([1, 16, 256])]; + tensor var_660_end_mask_0 = const()[name = tensor("op_660_end_mask_0"), val = tensor([true, true, true])]; + tensor var_660 = slice_by_index(begin = var_660_begin_0, end = var_660_end_0, end_mask = var_660_end_mask_0, x = window_23)[name = tensor("op_660")]; + tensor window_25_interleave_0 = const()[name = tensor("window_25_interleave_0"), val = tensor(false)]; + tensor window_25 = concat(axis = var_26, interleave = window_25_interleave_0, values = (var_660, var_657))[name = tensor("window_25")]; + tensor var_665_begin_0 = const()[name = tensor("op_665_begin_0"), val = tensor([0, 2, 0])]; + tensor var_665_end_0 = const()[name = tensor("op_665_end_0"), val = tensor([1, 3, 256])]; + tensor var_665_end_mask_0 = const()[name = tensor("op_665_end_mask_0"), val = tensor([true, false, true])]; + tensor var_665 = slice_by_index(begin = var_665_begin_0, end = var_665_end_0, end_mask = var_665_end_mask_0, x = x_15)[name = tensor("op_665")]; + tensor var_668_begin_0 = const()[name = tensor("op_668_begin_0"), val = tensor([0, 1, 0])]; + tensor var_668_end_0 = const()[name = tensor("op_668_end_0"), val = tensor([1, 16, 256])]; + tensor var_668_end_mask_0 = const()[name = tensor("op_668_end_mask_0"), val = tensor([true, true, true])]; + tensor var_668 = slice_by_index(begin = var_668_begin_0, end = var_668_end_0, end_mask = var_668_end_mask_0, x = window_25)[name = tensor("op_668")]; + tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; + tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_668, var_665))[name = tensor("window_27")]; + tensor var_673_begin_0 = const()[name = tensor("op_673_begin_0"), val = tensor([0, 3, 0])]; + tensor var_673_end_0 = const()[name = tensor("op_673_end_0"), val = tensor([1, 1, 256])]; + tensor var_673_end_mask_0 = const()[name = tensor("op_673_end_mask_0"), val = tensor([true, true, true])]; + tensor var_673 = slice_by_index(begin = var_673_begin_0, end = var_673_end_0, end_mask = var_673_end_mask_0, x = x_15)[name = tensor("op_673")]; + tensor var_676_begin_0 = const()[name = tensor("op_676_begin_0"), val = tensor([0, 1, 0])]; + tensor var_676_end_0 = const()[name = tensor("op_676_end_0"), val = tensor([1, 16, 256])]; + tensor var_676_end_mask_0 = const()[name = tensor("op_676_end_mask_0"), val = tensor([true, true, true])]; + tensor var_676 = slice_by_index(begin = var_676_begin_0, end = var_676_end_0, end_mask = var_676_end_mask_0, x = window_27)[name = tensor("op_676")]; + tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; + tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_676, var_673))[name = tensor("window_29")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_23, window_25, window_27, window_29))[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_701_split_sizes_0 = const()[name = tensor("op_701_split_sizes_0"), val = tensor([256, 256])]; + tensor var_701_axis_0 = const()[name = tensor("op_701_axis_0"), val = tensor(1)]; + tensor var_701_0, tensor var_701_1 = split(axis = var_701_axis_0, split_sizes = var_701_split_sizes_0, x = inputs_23)[name = tensor("op_701")]; + tensor var_703 = sigmoid(x = var_701_1)[name = tensor("op_703")]; + tensor inputs_25 = mul(x = var_701_0, y = var_703)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([4, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_734_begin_0 = const()[name = tensor("op_734_begin_0"), val = tensor([0, -1, 0])]; + tensor var_734_end_0 = const()[name = tensor("op_734_end_0"), val = tensor([4, 16, 256])]; + tensor var_734_end_mask_0 = const()[name = tensor("op_734_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_734 = slice_by_index(begin = var_734_begin_0, end = var_734_end_0, end_mask = var_734_end_mask_0, x = conv_out_5)[name = tensor("op_734")]; + tensor var_736_perm_0 = const()[name = tensor("op_736_perm_0"), val = tensor([1, 0, 2])]; + tensor var_736 = transpose(perm = var_736_perm_0, x = var_734)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_736)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_759 = const()[name = tensor("op_759"), val = tensor(0x1p-1)]; + tensor var_760 = mul(x = input_119, y = var_759)[name = tensor("op_760")]; + tensor input_121 = add(x = var_760, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_789 = const()[name = tensor("op_789"), val = tensor(0x1p-1)]; + tensor var_790 = mul(x = input_131, y = var_789)[name = tensor("op_790")]; + tensor input_133 = add(x = var_790, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_804 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 4, 4, 64])]; + tensor var_806 = reshape(shape = var_805, x = var_804)[name = tensor("op_806")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_810 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_811 = const()[name = tensor("op_811"), val = tensor(0x1p-3)]; + tensor var_812 = mul(x = var_810, y = var_811)[name = tensor("op_812")]; + tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 4, 4, 64])]; + tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_818 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_819 = const()[name = tensor("op_819"), val = tensor([1, 4, 4, 64])]; + tensor var_820 = reshape(shape = var_819, x = var_818)[name = tensor("op_820")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_814)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_806)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_830 = const()[name = tensor("op_830"), val = tensor([4, 1])]; + tensor var_831 = reshape(shape = var_830, x = sqrt_s_t_7)[name = tensor("op_831")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_831)[name = tensor("M_7")]; + tensor var_833 = mul(x = qk_7, y = M_7)[name = tensor("op_833")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_820)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_833, y = v_7)[name = tensor("inner_7")]; + tensor var_835_transpose_x_0 = const()[name = tensor("op_835_transpose_x_0"), val = tensor(false)]; + tensor var_835_transpose_y_0 = const()[name = tensor("op_835_transpose_y_0"), val = tensor(false)]; + tensor var_835 = matmul(transpose_x = var_835_transpose_x_0, transpose_y = var_835_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_835")]; + tensor var_836 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_836")]; + tensor var_837 = const()[name = tensor("op_837"), val = tensor([1, 1, 4, 1])]; + tensor var_838 = reshape(shape = var_837, x = var_836)[name = tensor("op_838")]; + tensor cross_7 = mul(x = var_835, y = var_838)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_841 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_841")]; + tensor var_843_transpose_x_1 = const()[name = tensor("op_843_transpose_x_1"), val = tensor(true)]; + tensor var_843_transpose_y_1 = const()[name = tensor("op_843_transpose_y_1"), val = tensor(false)]; + tensor var_843 = matmul(transpose_x = var_843_transpose_x_1, transpose_y = var_843_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_843")]; + tensor new_kv_unnorm_7 = add(x = var_841, y = var_843)[name = tensor("new_kv_unnorm_7")]; + tensor var_845 = const()[name = tensor("op_845"), val = tensor(0x1p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_845)[name = tensor("new_scale_7")]; + tensor var_847 = sqrt(x = new_scale_7)[name = tensor("op_847")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_847)[name = tensor("nkv_1")]; + tensor var_849_perm_0 = const()[name = tensor("op_849_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_849 = transpose(perm = var_849_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_849)[name = tensor("out_21")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 4, 256])]; + tensor out_23 = reshape(shape = var_853, x = out_21)[name = tensor("out_23")]; + tensor var_855 = silu(x = input_137)[name = tensor("op_855")]; + tensor input_139 = mul(x = var_855, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_31_begin_0 = const()[name = tensor("window_31_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_31_end_0 = const()[name = tensor("window_31_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_31_end_mask_0 = const()[name = tensor("window_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_31_squeeze_mask_0 = const()[name = tensor("window_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_31 = slice_by_index(begin = window_31_begin_0, end = window_31_end_0, end_mask = window_31_end_mask_0, squeeze_mask = window_31_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_31")]; + tensor var_863_begin_0 = const()[name = tensor("op_863_begin_0"), val = tensor([0, 0, 0])]; + tensor var_863_end_0 = const()[name = tensor("op_863_end_0"), val = tensor([1, 1, 256])]; + tensor var_863_end_mask_0 = const()[name = tensor("op_863_end_mask_0"), val = tensor([true, false, true])]; + tensor var_863 = slice_by_index(begin = var_863_begin_0, end = var_863_end_0, end_mask = var_863_end_mask_0, x = x_21)[name = tensor("op_863")]; + tensor var_866_begin_0 = const()[name = tensor("op_866_begin_0"), val = tensor([0, 1, 0])]; + tensor var_866_end_0 = const()[name = tensor("op_866_end_0"), val = tensor([1, 16, 256])]; + tensor var_866_end_mask_0 = const()[name = tensor("op_866_end_mask_0"), val = tensor([true, true, true])]; + tensor var_866 = slice_by_index(begin = var_866_begin_0, end = var_866_end_0, end_mask = var_866_end_mask_0, x = window_31)[name = tensor("op_866")]; + tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; + tensor window_33 = concat(axis = var_26, interleave = window_33_interleave_0, values = (var_866, var_863))[name = tensor("window_33")]; + tensor var_871_begin_0 = const()[name = tensor("op_871_begin_0"), val = tensor([0, 1, 0])]; + tensor var_871_end_0 = const()[name = tensor("op_871_end_0"), val = tensor([1, 2, 256])]; + tensor var_871_end_mask_0 = const()[name = tensor("op_871_end_mask_0"), val = tensor([true, false, true])]; + tensor var_871 = slice_by_index(begin = var_871_begin_0, end = var_871_end_0, end_mask = var_871_end_mask_0, x = x_21)[name = tensor("op_871")]; + tensor var_874_begin_0 = const()[name = tensor("op_874_begin_0"), val = tensor([0, 1, 0])]; + tensor var_874_end_0 = const()[name = tensor("op_874_end_0"), val = tensor([1, 16, 256])]; + tensor var_874_end_mask_0 = const()[name = tensor("op_874_end_mask_0"), val = tensor([true, true, true])]; + tensor var_874 = slice_by_index(begin = var_874_begin_0, end = var_874_end_0, end_mask = var_874_end_mask_0, x = window_33)[name = tensor("op_874")]; + tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; + tensor window_35 = concat(axis = var_26, interleave = window_35_interleave_0, values = (var_874, var_871))[name = tensor("window_35")]; + tensor var_879_begin_0 = const()[name = tensor("op_879_begin_0"), val = tensor([0, 2, 0])]; + tensor var_879_end_0 = const()[name = tensor("op_879_end_0"), val = tensor([1, 3, 256])]; + tensor var_879_end_mask_0 = const()[name = tensor("op_879_end_mask_0"), val = tensor([true, false, true])]; + tensor var_879 = slice_by_index(begin = var_879_begin_0, end = var_879_end_0, end_mask = var_879_end_mask_0, x = x_21)[name = tensor("op_879")]; + tensor var_882_begin_0 = const()[name = tensor("op_882_begin_0"), val = tensor([0, 1, 0])]; + tensor var_882_end_0 = const()[name = tensor("op_882_end_0"), val = tensor([1, 16, 256])]; + tensor var_882_end_mask_0 = const()[name = tensor("op_882_end_mask_0"), val = tensor([true, true, true])]; + tensor var_882 = slice_by_index(begin = var_882_begin_0, end = var_882_end_0, end_mask = var_882_end_mask_0, x = window_35)[name = tensor("op_882")]; + tensor window_37_interleave_0 = const()[name = tensor("window_37_interleave_0"), val = tensor(false)]; + tensor window_37 = concat(axis = var_26, interleave = window_37_interleave_0, values = (var_882, var_879))[name = tensor("window_37")]; + tensor var_887_begin_0 = const()[name = tensor("op_887_begin_0"), val = tensor([0, 3, 0])]; + tensor var_887_end_0 = const()[name = tensor("op_887_end_0"), val = tensor([1, 1, 256])]; + tensor var_887_end_mask_0 = const()[name = tensor("op_887_end_mask_0"), val = tensor([true, true, true])]; + tensor var_887 = slice_by_index(begin = var_887_begin_0, end = var_887_end_0, end_mask = var_887_end_mask_0, x = x_21)[name = tensor("op_887")]; + tensor var_890_begin_0 = const()[name = tensor("op_890_begin_0"), val = tensor([0, 1, 0])]; + tensor var_890_end_0 = const()[name = tensor("op_890_end_0"), val = tensor([1, 16, 256])]; + tensor var_890_end_mask_0 = const()[name = tensor("op_890_end_mask_0"), val = tensor([true, true, true])]; + tensor var_890 = slice_by_index(begin = var_890_begin_0, end = var_890_end_0, end_mask = var_890_end_mask_0, x = window_37)[name = tensor("op_890")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_890, var_887))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_33, window_35, window_37, window))[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_915_split_sizes_0 = const()[name = tensor("op_915_split_sizes_0"), val = tensor([256, 256])]; + tensor var_915_axis_0 = const()[name = tensor("op_915_axis_0"), val = tensor(1)]; + tensor var_915_0, tensor var_915_1 = split(axis = var_915_axis_0, split_sizes = var_915_split_sizes_0, x = inputs_33)[name = tensor("op_915")]; + tensor var_917 = sigmoid(x = var_915_1)[name = tensor("op_917")]; + tensor inputs_35 = mul(x = var_915_0, y = var_917)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([4, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_948_begin_0 = const()[name = tensor("op_948_begin_0"), val = tensor([0, -1, 0])]; + tensor var_948_end_0 = const()[name = tensor("op_948_end_0"), val = tensor([4, 16, 256])]; + tensor var_948_end_mask_0 = const()[name = tensor("op_948_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_948 = slice_by_index(begin = var_948_begin_0, end = var_948_end_0, end_mask = var_948_end_mask_0, x = conv_out_7)[name = tensor("op_948")]; + tensor var_950_perm_0 = const()[name = tensor("op_950_perm_0"), val = tensor([1, 0, 2])]; + tensor var_950 = transpose(perm = var_950_perm_0, x = var_948)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_950)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_973 = const()[name = tensor("op_973"), val = tensor(0x1p-1)]; + tensor var_974 = mul(x = input_159, y = var_973)[name = tensor("op_974")]; + tensor input_161 = add(x = var_974, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_992_begin_0 = const()[name = tensor("op_992_begin_0"), val = tensor([0, 0, 4])]; + tensor var_992_end_0 = const()[name = tensor("op_992_end_0"), val = tensor([1, 256, 22])]; + tensor var_992_end_mask_0 = const()[name = tensor("op_992_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_992_begin_0, end = var_992_end_0, end_mask = var_992_end_mask_0, x = cat)[name = tensor("op_992")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_994 = const()[name = tensor("op_994"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_995 = reduce_l2_norm(axes = var_994, keep_dims = var_29, x = input_163)[name = tensor("op_995")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_995)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_999_axis_0 = const()[name = tensor("op_999_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_999_axis_0, values = (var_206, var_420, var_634, nkv_1))[name = tensor("op_999")]; + tensor var_1001_axis_0 = const()[name = tensor("op_1001_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1001_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1001")]; + tensor var_1003_axis_0 = const()[name = tensor("op_1003_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1003_axis_0, values = (window_9, window_19, window_29, window))[name = tensor("op_1003")]; + tensor var_1012 = const()[name = tensor("op_1012"), val = tensor(0x1.5798eep-27)]; + tensor var_1017 = const()[name = tensor("op_1017"), val = tensor(0x1.4f8b58p-17)]; + tensor var_1019 = const()[name = tensor("op_1019"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_1020 = const()[name = tensor("op_1020"), val = tensor(true)]; + tensor var_1022 = const()[name = tensor("op_1022"), val = tensor(0x1p+0)]; + tensor var_1026 = const()[name = tensor("op_1026"), val = tensor(-1)]; + tensor var_1032 = const()[name = tensor("op_1032"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395584)))]; + tensor var_1094_axes_0 = const()[name = tensor("op_1094_axes_0"), val = tensor([2])]; + tensor var_1094 = expand_dims(axes = var_1094_axes_0, x = emb)[name = tensor("op_1094")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1094)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_1026, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1102_perm_0 = const()[name = tensor("op_1102_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([12, 4, 256])]; + tensor var_1102 = transpose(perm = var_1102_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1106, x = var_1102)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1114 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1115 = const()[name = tensor("op_1115"), val = tensor([12, 4, 4, 64])]; + tensor var_1116 = reshape(shape = var_1115, x = var_1114)[name = tensor("op_1116")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1120 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1121 = const()[name = tensor("op_1121"), val = tensor(0x1p-3)]; + tensor var_1122 = mul(x = var_1120, y = var_1121)[name = tensor("op_1122")]; + tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([12, 4, 4, 64])]; + tensor var_1124 = reshape(shape = var_1123, x = var_1122)[name = tensor("op_1124")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1128 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1129 = const()[name = tensor("op_1129"), val = tensor([12, 4, 4, 64])]; + tensor var_1130 = reshape(shape = var_1129, x = var_1128)[name = tensor("op_1130")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_1032, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_1022, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1124)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1116)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([1, 4])]; + tensor var_1143 = reshape(shape = var_1142, x = valid_mask)[name = tensor("op_1143")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1143)[name = tensor("causal_with_valid_1")]; + tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([4, 1])]; + tensor var_1146 = reshape(shape = var_1145, x = sqrt_s_t_9)[name = tensor("op_1146")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1146)[name = tensor("M_9")]; + tensor var_1148 = mul(x = qk_9, y = M_9)[name = tensor("op_1148")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1130)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1148, y = v_9)[name = tensor("inner_9")]; + tensor var_1150_transpose_x_0 = const()[name = tensor("op_1150_transpose_x_0"), val = tensor(false)]; + tensor var_1150_transpose_y_0 = const()[name = tensor("op_1150_transpose_y_0"), val = tensor(false)]; + tensor var_1150 = matmul(transpose_x = var_1150_transpose_x_0, transpose_y = var_1150_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1150")]; + tensor var_1151 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1151")]; + tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 1, 4, 1])]; + tensor var_1153 = reshape(shape = var_1152, x = var_1151)[name = tensor("op_1153")]; + tensor cross_9 = mul(x = var_1150, y = var_1153)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 1, 4, 1])]; + tensor var_1157 = reshape(shape = var_1156, x = valid_mask)[name = tensor("op_1157")]; + tensor v_masked_1 = mul(x = v_9, y = var_1157)[name = tensor("v_masked_1")]; + tensor var_1159 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1159")]; + tensor var_1161_transpose_x_1 = const()[name = tensor("op_1161_transpose_x_1"), val = tensor(true)]; + tensor var_1161_transpose_y_1 = const()[name = tensor("op_1161_transpose_y_1"), val = tensor(false)]; + tensor var_1161 = matmul(transpose_x = var_1161_transpose_x_1, transpose_y = var_1161_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1161")]; + tensor new_kv_unnorm_9 = add(x = var_1159, y = var_1161)[name = tensor("new_kv_unnorm_9")]; + tensor var_1163_keep_dims_0 = const()[name = tensor("op_1163_keep_dims_0"), val = tensor(false)]; + tensor var_1163 = reduce_sum(keep_dims = var_1163_keep_dims_0, x = valid_mask)[name = tensor("op_1163")]; + tensor var_1164 = const()[name = tensor("op_1164"), val = tensor([1])]; + tensor var_1165 = reshape(shape = var_1164, x = var_1163)[name = tensor("op_1165")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1165)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_1022, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1169 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1169")]; + tensor var_1170_perm_0 = const()[name = tensor("op_1170_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1170 = transpose(perm = var_1170_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1019, x = var_1170)[name = tensor("out_27")]; + tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([12, 4, 256])]; + tensor out_29 = reshape(shape = var_1174, x = out_27)[name = tensor("out_29")]; + tensor var_1176 = silu(x = input_169)[name = tensor("op_1176")]; + tensor input_171 = mul(x = var_1176, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1017, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1186 = const()[name = tensor("op_1186"), val = tensor([1, 12, 4, 256])]; + tensor var_1187 = reshape(shape = var_1186, x = xt_1)[name = tensor("op_1187")]; + tensor var_1188_perm_0 = const()[name = tensor("op_1188_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([4, 12, 256])]; + tensor var_1188 = transpose(perm = var_1188_perm_0, x = var_1187)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1191, x = var_1188)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1214 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 4, 3, 256])]; + tensor var_1216 = reshape(shape = concat_1, x = var_1214)[name = tensor("op_1216")]; + tensor var_1217_axes_0 = const()[name = tensor("op_1217_axes_0"), val = tensor([0])]; + tensor var_1217 = expand_dims(axes = var_1217_axes_0, x = var_1216)[name = tensor("op_1217")]; + tensor var_1218_perm_0 = const()[name = tensor("op_1218_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1219_axes_0 = const()[name = tensor("op_1219_axes_0"), val = tensor([-2])]; + tensor var_1218 = transpose(perm = var_1218_perm_0, x = var_1217)[name = tensor("transpose_21")]; + tensor var_1219 = squeeze(axes = var_1219_axes_0, x = var_1218)[name = tensor("op_1219")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 4, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1219)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 4, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1219)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 4, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1219)[name = tensor("v_11")]; + tensor var_1227 = const()[name = tensor("op_1227"), val = tensor([12, 16, 64])]; + tensor var_1228 = reshape(shape = var_1227, x = q_11)[name = tensor("op_1228")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1234 = const()[name = tensor("op_1234"), val = tensor([12, 16, 64])]; + tensor var_1235 = reshape(shape = var_1234, x = k_11)[name = tensor("op_1235")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1241 = const()[name = tensor("op_1241"), val = tensor([12, 16, 64])]; + tensor var_1242 = reshape(shape = var_1241, x = v_11)[name = tensor("op_1242")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([4, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1228)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1245, x = q_13)[name = tensor("q_15")]; + tensor var_1247 = const()[name = tensor("op_1247"), val = tensor([4, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1235)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1247, x = k_13)[name = tensor("k_15")]; + tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([4, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1242)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1249, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([2, 0, 1, 3])]; + tensor var_1257 = const()[name = tensor("op_1257"), val = tensor([48, 256])]; + tensor var_1253 = transpose(perm = var_1252, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1257, x = var_1253)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([12, 4, 256])]; + tensor attn_output_7 = reshape(shape = var_1261, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1017, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1017, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1281 = const()[name = tensor("op_1281"), val = tensor([1, 4, 12, 256])]; + tensor x_31 = reshape(shape = var_1281, x = xt_3)[name = tensor("x_31")]; + tensor var_1283_perm_0 = const()[name = tensor("op_1283_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1287 = const()[name = tensor("op_1287"), val = tensor([12, 4, 256])]; + tensor var_1283 = transpose(perm = var_1283_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1287, x = var_1283)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1295 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1296 = const()[name = tensor("op_1296"), val = tensor([12, 4, 4, 64])]; + tensor var_1297 = reshape(shape = var_1296, x = var_1295)[name = tensor("op_1297")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1301 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1302 = const()[name = tensor("op_1302"), val = tensor(0x1p-3)]; + tensor var_1303 = mul(x = var_1301, y = var_1302)[name = tensor("op_1303")]; + tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([12, 4, 4, 64])]; + tensor var_1305 = reshape(shape = var_1304, x = var_1303)[name = tensor("op_1305")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1309 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1310 = const()[name = tensor("op_1310"), val = tensor([12, 4, 4, 64])]; + tensor var_1311 = reshape(shape = var_1310, x = var_1309)[name = tensor("op_1311")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_1022, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1305)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1297)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([4, 1])]; + tensor var_1327 = reshape(shape = var_1326, x = sqrt_s_t)[name = tensor("op_1327")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1327)[name = tensor("M")]; + tensor var_1329 = mul(x = qk, y = M)[name = tensor("op_1329")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1311)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1329, y = v_17)[name = tensor("inner")]; + tensor var_1331_transpose_x_0 = const()[name = tensor("op_1331_transpose_x_0"), val = tensor(false)]; + tensor var_1331_transpose_y_0 = const()[name = tensor("op_1331_transpose_y_0"), val = tensor(false)]; + tensor var_1331 = matmul(transpose_x = var_1331_transpose_x_0, transpose_y = var_1331_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1331")]; + tensor var_1332 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1332")]; + tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 1, 4, 1])]; + tensor var_1334 = reshape(shape = var_1333, x = var_1332)[name = tensor("op_1334")]; + tensor cross = mul(x = var_1331, y = var_1334)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1157)[name = tensor("v_masked")]; + tensor var_1340 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1340")]; + tensor var_1342_transpose_x_1 = const()[name = tensor("op_1342_transpose_x_1"), val = tensor(true)]; + tensor var_1342_transpose_y_1 = const()[name = tensor("op_1342_transpose_y_1"), val = tensor(false)]; + tensor var_1342 = matmul(transpose_x = var_1342_transpose_x_1, transpose_y = var_1342_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1342")]; + tensor new_kv_unnorm = add(x = var_1340, y = var_1342)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1165)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_1022, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1351_perm_0 = const()[name = tensor("op_1351_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1351 = transpose(perm = var_1351_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1019, x = var_1351)[name = tensor("out_33")]; + tensor var_1355 = const()[name = tensor("op_1355"), val = tensor([12, 4, 256])]; + tensor out = reshape(shape = var_1355, x = out_33)[name = tensor("out")]; + tensor var_1357 = silu(x = input_187)[name = tensor("op_1357")]; + tensor input_189 = mul(x = var_1357, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1017, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1367 = const()[name = tensor("op_1367"), val = tensor([1, 12, 4, 256])]; + tensor var_1368 = reshape(shape = var_1367, x = xt_5)[name = tensor("op_1368")]; + tensor var_1369_perm_0 = const()[name = tensor("op_1369_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1372 = const()[name = tensor("op_1372"), val = tensor([4, 12, 256])]; + tensor var_1369 = transpose(perm = var_1369_perm_0, x = var_1368)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1372, x = var_1369)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1395 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 4, 3, 256])]; + tensor var_1397 = reshape(shape = concat_2, x = var_1395)[name = tensor("op_1397")]; + tensor var_1398_axes_0 = const()[name = tensor("op_1398_axes_0"), val = tensor([0])]; + tensor var_1398 = expand_dims(axes = var_1398_axes_0, x = var_1397)[name = tensor("op_1398")]; + tensor var_1399_perm_0 = const()[name = tensor("op_1399_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1400_axes_0 = const()[name = tensor("op_1400_axes_0"), val = tensor([-2])]; + tensor var_1399 = transpose(perm = var_1399_perm_0, x = var_1398)[name = tensor("transpose_8")]; + tensor var_1400 = squeeze(axes = var_1400_axes_0, x = var_1399)[name = tensor("op_1400")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 4, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1400)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 4, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1400)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 4, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1400)[name = tensor("v_19")]; + tensor var_1408 = const()[name = tensor("op_1408"), val = tensor([12, 16, 64])]; + tensor var_1409 = reshape(shape = var_1408, x = q_19)[name = tensor("op_1409")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1415 = const()[name = tensor("op_1415"), val = tensor([12, 16, 64])]; + tensor var_1416 = reshape(shape = var_1415, x = k_19)[name = tensor("op_1416")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1422 = const()[name = tensor("op_1422"), val = tensor([12, 16, 64])]; + tensor var_1423 = reshape(shape = var_1422, x = v_19)[name = tensor("op_1423")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1426 = const()[name = tensor("op_1426"), val = tensor([4, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1409)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1426, x = q_21)[name = tensor("q")]; + tensor var_1428 = const()[name = tensor("op_1428"), val = tensor([4, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1416)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1428, x = k_21)[name = tensor("k")]; + tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([4, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1423)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1430, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([2, 0, 1, 3])]; + tensor var_1438 = const()[name = tensor("op_1438"), val = tensor([48, 256])]; + tensor var_1434 = transpose(perm = var_1433, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1438, x = var_1434)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([12, 4, 256])]; + tensor attn_output = reshape(shape = var_1442, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1017, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1017, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1462 = const()[name = tensor("op_1462"), val = tensor([1, 4, 12, 256])]; + tensor input = reshape(shape = var_1462, x = xt)[name = tensor("input")]; + tensor var_1464 = const()[name = tensor("op_1464"), val = tensor([-1])]; + tensor var_1465 = reduce_l2_norm(axes = var_1464, keep_dims = var_1020, x = input)[name = tensor("op_1465")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_1012, beta = const_42, x = var_1465)[name = tensor("clip_5")]; + tensor var_1467 = real_div(x = input, y = clip_5)[name = tensor("op_1467")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([4, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([4, 256, 12])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1467)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 4, 12])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 4, 11])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1471")]; + tensor var_1473_axis_0 = const()[name = tensor("op_1473_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1473_axis_0, values = (var_1169, nkv))[name = tensor("op_1473")]; + tensor var_1475_axis_0 = const()[name = tensor("op_1475_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1475_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1475")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/weights/weight.bin b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..059cd18c26262ebe0f0801492548e073ef051285 --- /dev/null +++ b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:293b8809fc4de66da4495caa66b6814a3764ac8702bedec17612b1194542622f +size 44444800 diff --git a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..8c1ca861fb9558eb088d1aeb653c17845f7087b0 --- /dev/null +++ b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:512460ea937729458d31e1520a4866d8d4204db686d08c065f181bcfa9f13c79 +size 191051 diff --git a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..059cd18c26262ebe0f0801492548e073ef051285 --- /dev/null +++ b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:293b8809fc4de66da4495caa66b6814a3764ac8702bedec17612b1194542622f +size 44444800 diff --git a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Manifest.json b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..1f2e5792778d89ff02951ec657bbd30348216ac5 --- /dev/null +++ b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "8AAD4EFD-A1AE-4676-AFDF-014CF1A0D1F3": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "F3FC4090-5689-4260-9E54-5F399C74FF28": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "8AAD4EFD-A1AE-4676-AFDF-014CF1A0D1F3" +} diff --git a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/analytics/coremldata.bin b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..71b915acdd9abe5171728dadc38af3b7263e0247 --- /dev/null +++ b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3ab2696bac4217bb0bfc9f74776b079447b45424191c9523d5d87f9a001c761 +size 243 diff --git a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/coremldata.bin b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..f1fe815e519cef7d8e571c8c3bb4fc3c51bf16bd --- /dev/null +++ b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8aa47228de8673fd3089ab1dcf43fd425952e0eeaf3be2efeb7db0bdae626a6 +size 1308 diff --git a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/metadata.json b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..118719eec347c4d50bae0ae8b401d770b4bf1ac8 --- /dev/null +++ b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND DIHARD II streaming diarizer (pipeline, T=5, max_speakers=10)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 5 × 10)", + "shortDescription" : "", + "shape" : "[1, 5, 10]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 68, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 72, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 26, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 47 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 5 × 345)", + "shortDescription" : "", + "shape" : "[1, 5, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 5)", + "shortDescription" : "", + "shape" : "[5]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"dih2\", \"model_label\": \"DIHARD II\", \"variant\": \"pipeline\", \"chunk_size\": 5, \"step_duration_ms\": 500, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_dih2_500ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/model.mil b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..d06d8cb5807c4c0b54c3e07f38ba5655c925422a --- /dev/null +++ b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/model.mil @@ -0,0 +1,1366 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2, 0x1.4p+2])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982144)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983232)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336576)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337664)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338752)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339840)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340928)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345088)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393728)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394816)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443456)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444544)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445632)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446720)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708928)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7710016)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972224)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973312)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235520)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236608)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498816)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499904)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762112)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763200)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764288)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766400)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290752)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307200)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308288)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309376)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310464)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311552)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312640)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574848)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575936)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9577024)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581184)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629824)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630912)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679552)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680640)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681728)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682816)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683904)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688064)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736704)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737792)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786432)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787520)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788608)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789696)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051904)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052992)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315200)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316288)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578496)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579584)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841792)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842880)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105088)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106176)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107264)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109376)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633728)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650176)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651264)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652352)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653440)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654528)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655616)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917824)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918912)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15920000)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924160)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972800)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973888)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022528)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023616)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024704)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025792)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026880)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18031040)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079680)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080768)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129408)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130496)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131584)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132672)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394880)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395968)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658176)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659264)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921472)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922560)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184768)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185856)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448064)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449152)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450240)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452352)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976704)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993152)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994240)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995328)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996416)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997504)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998592)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260800)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261888)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262976)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267136)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315776)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316864)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365504)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366592)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367680)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368768)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369856)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24374016)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422656)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423744)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472384)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473472)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474560)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475648)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737856)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738944)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001152)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002240)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264448)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265536)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527744)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528832)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791040)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792128)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793216)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795328)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319680)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336128)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337216)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338304)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339392)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340480)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341568)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603776)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604864)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605952)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610112)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658752)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659840)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708480)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709568)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710656)))]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710848)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711936)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236288)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237376)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499584)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500672)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762880)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763968)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026176)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027264)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289472)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290560)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552768)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553856)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554944)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32556032)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818240)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821376)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607872)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608960)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33610048)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618304)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715520)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716608)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813824)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814912)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816000)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37817088)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079296)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080384)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342592)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343680)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605888)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606976)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869184)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870272)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132480)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133568)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134656)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135744)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397952)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39401088)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187584)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188672)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189760)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40198016)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295232)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296320)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393536)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394624)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_18 = const()[name = tensor("op_18"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_21 = const()[name = tensor("op_21"), val = tensor(2)]; + tensor var_24 = const()[name = tensor("op_24"), val = tensor(0)]; + tensor var_27 = const()[name = tensor("op_27"), val = tensor(1)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(0x1.4f8b58p-17)]; + tensor var_30 = const()[name = tensor("op_30"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_29, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_148 = const()[name = tensor("op_148"), val = tensor(0x1p-1)]; + tensor var_149 = mul(x = input_11, y = var_148)[name = tensor("op_149")]; + tensor input_13 = add(x = var_149, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_29, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_163 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_164 = const()[name = tensor("op_164"), val = tensor([1, 5, 4, 64])]; + tensor var_165 = reshape(shape = var_164, x = var_163)[name = tensor("op_165")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_169 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_170 = const()[name = tensor("op_170"), val = tensor(0x1p-3)]; + tensor var_171 = mul(x = var_169, y = var_170)[name = tensor("op_171")]; + tensor var_172 = const()[name = tensor("op_172"), val = tensor([1, 5, 4, 64])]; + tensor var_173 = reshape(shape = var_172, x = var_171)[name = tensor("op_173")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_177 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_178 = const()[name = tensor("op_178"), val = tensor([1, 5, 4, 64])]; + tensor var_179 = reshape(shape = var_178, x = var_177)[name = tensor("op_179")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_173)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_165)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_189 = const()[name = tensor("op_189"), val = tensor([5, 1])]; + tensor var_190 = reshape(shape = var_189, x = sqrt_s_t_1)[name = tensor("op_190")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_190)[name = tensor("M_1")]; + tensor var_192 = mul(x = qk_1, y = M_1)[name = tensor("op_192")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_179)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_192, y = v_1)[name = tensor("inner_1")]; + tensor var_194_transpose_x_0 = const()[name = tensor("op_194_transpose_x_0"), val = tensor(false)]; + tensor var_194_transpose_y_0 = const()[name = tensor("op_194_transpose_y_0"), val = tensor(false)]; + tensor var_194 = matmul(transpose_x = var_194_transpose_x_0, transpose_y = var_194_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_194")]; + tensor var_195 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_195")]; + tensor var_196 = const()[name = tensor("op_196"), val = tensor([1, 1, 5, 1])]; + tensor var_197 = reshape(shape = var_196, x = var_195)[name = tensor("op_197")]; + tensor cross_1 = mul(x = var_194, y = var_197)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_200 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_200")]; + tensor var_202_transpose_x_1 = const()[name = tensor("op_202_transpose_x_1"), val = tensor(true)]; + tensor var_202_transpose_y_1 = const()[name = tensor("op_202_transpose_y_1"), val = tensor(false)]; + tensor var_202 = matmul(transpose_x = var_202_transpose_x_1, transpose_y = var_202_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_202")]; + tensor new_kv_unnorm_1 = add(x = var_200, y = var_202)[name = tensor("new_kv_unnorm_1")]; + tensor var_204 = const()[name = tensor("op_204"), val = tensor(0x1.4p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_204)[name = tensor("new_scale_1")]; + tensor var_206 = sqrt(x = new_scale_1)[name = tensor("op_206")]; + tensor var_207 = real_div(x = new_kv_unnorm_1, y = var_206)[name = tensor("op_207")]; + tensor var_208_perm_0 = const()[name = tensor("op_208_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_208 = transpose(perm = var_208_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_18, x = var_208)[name = tensor("out_3")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor([1, 5, 256])]; + tensor out_5 = reshape(shape = var_212, x = out_3)[name = tensor("out_5")]; + tensor var_214 = silu(x = input_17)[name = tensor("op_214")]; + tensor input_19 = mul(x = var_214, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_222_begin_0 = const()[name = tensor("op_222_begin_0"), val = tensor([0, 0, 0])]; + tensor var_222_end_0 = const()[name = tensor("op_222_end_0"), val = tensor([1, 1, 256])]; + tensor var_222_end_mask_0 = const()[name = tensor("op_222_end_mask_0"), val = tensor([true, false, true])]; + tensor var_222 = slice_by_index(begin = var_222_begin_0, end = var_222_end_0, end_mask = var_222_end_mask_0, x = x_3)[name = tensor("op_222")]; + tensor var_225_begin_0 = const()[name = tensor("op_225_begin_0"), val = tensor([0, 1, 0])]; + tensor var_225_end_0 = const()[name = tensor("op_225_end_0"), val = tensor([1, 16, 256])]; + tensor var_225_end_mask_0 = const()[name = tensor("op_225_end_mask_0"), val = tensor([true, true, true])]; + tensor var_225 = slice_by_index(begin = var_225_begin_0, end = var_225_end_0, end_mask = var_225_end_mask_0, x = window_1)[name = tensor("op_225")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_27, interleave = window_3_interleave_0, values = (var_225, var_222))[name = tensor("window_3")]; + tensor var_230_begin_0 = const()[name = tensor("op_230_begin_0"), val = tensor([0, 1, 0])]; + tensor var_230_end_0 = const()[name = tensor("op_230_end_0"), val = tensor([1, 2, 256])]; + tensor var_230_end_mask_0 = const()[name = tensor("op_230_end_mask_0"), val = tensor([true, false, true])]; + tensor var_230 = slice_by_index(begin = var_230_begin_0, end = var_230_end_0, end_mask = var_230_end_mask_0, x = x_3)[name = tensor("op_230")]; + tensor var_233_begin_0 = const()[name = tensor("op_233_begin_0"), val = tensor([0, 1, 0])]; + tensor var_233_end_0 = const()[name = tensor("op_233_end_0"), val = tensor([1, 16, 256])]; + tensor var_233_end_mask_0 = const()[name = tensor("op_233_end_mask_0"), val = tensor([true, true, true])]; + tensor var_233 = slice_by_index(begin = var_233_begin_0, end = var_233_end_0, end_mask = var_233_end_mask_0, x = window_3)[name = tensor("op_233")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_27, interleave = window_5_interleave_0, values = (var_233, var_230))[name = tensor("window_5")]; + tensor var_238_begin_0 = const()[name = tensor("op_238_begin_0"), val = tensor([0, 2, 0])]; + tensor var_238_end_0 = const()[name = tensor("op_238_end_0"), val = tensor([1, 3, 256])]; + tensor var_238_end_mask_0 = const()[name = tensor("op_238_end_mask_0"), val = tensor([true, false, true])]; + tensor var_238 = slice_by_index(begin = var_238_begin_0, end = var_238_end_0, end_mask = var_238_end_mask_0, x = x_3)[name = tensor("op_238")]; + tensor var_241_begin_0 = const()[name = tensor("op_241_begin_0"), val = tensor([0, 1, 0])]; + tensor var_241_end_0 = const()[name = tensor("op_241_end_0"), val = tensor([1, 16, 256])]; + tensor var_241_end_mask_0 = const()[name = tensor("op_241_end_mask_0"), val = tensor([true, true, true])]; + tensor var_241 = slice_by_index(begin = var_241_begin_0, end = var_241_end_0, end_mask = var_241_end_mask_0, x = window_5)[name = tensor("op_241")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_27, interleave = window_7_interleave_0, values = (var_241, var_238))[name = tensor("window_7")]; + tensor var_246_begin_0 = const()[name = tensor("op_246_begin_0"), val = tensor([0, 3, 0])]; + tensor var_246_end_0 = const()[name = tensor("op_246_end_0"), val = tensor([1, 4, 256])]; + tensor var_246_end_mask_0 = const()[name = tensor("op_246_end_mask_0"), val = tensor([true, false, true])]; + tensor var_246 = slice_by_index(begin = var_246_begin_0, end = var_246_end_0, end_mask = var_246_end_mask_0, x = x_3)[name = tensor("op_246")]; + tensor var_249_begin_0 = const()[name = tensor("op_249_begin_0"), val = tensor([0, 1, 0])]; + tensor var_249_end_0 = const()[name = tensor("op_249_end_0"), val = tensor([1, 16, 256])]; + tensor var_249_end_mask_0 = const()[name = tensor("op_249_end_mask_0"), val = tensor([true, true, true])]; + tensor var_249 = slice_by_index(begin = var_249_begin_0, end = var_249_end_0, end_mask = var_249_end_mask_0, x = window_7)[name = tensor("op_249")]; + tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; + tensor window_9 = concat(axis = var_27, interleave = window_9_interleave_0, values = (var_249, var_246))[name = tensor("window_9")]; + tensor var_254_begin_0 = const()[name = tensor("op_254_begin_0"), val = tensor([0, 4, 0])]; + tensor var_254_end_0 = const()[name = tensor("op_254_end_0"), val = tensor([1, 1, 256])]; + tensor var_254_end_mask_0 = const()[name = tensor("op_254_end_mask_0"), val = tensor([true, true, true])]; + tensor var_254 = slice_by_index(begin = var_254_begin_0, end = var_254_end_0, end_mask = var_254_end_mask_0, x = x_3)[name = tensor("op_254")]; + tensor var_257_begin_0 = const()[name = tensor("op_257_begin_0"), val = tensor([0, 1, 0])]; + tensor var_257_end_0 = const()[name = tensor("op_257_end_0"), val = tensor([1, 16, 256])]; + tensor var_257_end_mask_0 = const()[name = tensor("op_257_end_mask_0"), val = tensor([true, true, true])]; + tensor var_257 = slice_by_index(begin = var_257_begin_0, end = var_257_end_0, end_mask = var_257_end_mask_0, x = window_9)[name = tensor("op_257")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_27, interleave = window_11_interleave_0, values = (var_257, var_254))[name = tensor("window_11")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_24, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9, window_11))[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_282_split_sizes_0 = const()[name = tensor("op_282_split_sizes_0"), val = tensor([256, 256])]; + tensor var_282_axis_0 = const()[name = tensor("op_282_axis_0"), val = tensor(1)]; + tensor var_282_0, tensor var_282_1 = split(axis = var_282_axis_0, split_sizes = var_282_split_sizes_0, x = inputs_3)[name = tensor("op_282")]; + tensor var_284 = sigmoid(x = var_282_1)[name = tensor("op_284")]; + tensor inputs_5 = mul(x = var_282_0, y = var_284)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([5, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_315_begin_0 = const()[name = tensor("op_315_begin_0"), val = tensor([0, -1, 0])]; + tensor var_315_end_0 = const()[name = tensor("op_315_end_0"), val = tensor([5, 16, 256])]; + tensor var_315_end_mask_0 = const()[name = tensor("op_315_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_315 = slice_by_index(begin = var_315_begin_0, end = var_315_end_0, end_mask = var_315_end_mask_0, x = conv_out_1)[name = tensor("op_315")]; + tensor var_317_perm_0 = const()[name = tensor("op_317_perm_0"), val = tensor([1, 0, 2])]; + tensor var_317 = transpose(perm = var_317_perm_0, x = var_315)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_317)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor(0x1p-1)]; + tensor var_341 = mul(x = input_39, y = var_340)[name = tensor("op_341")]; + tensor input_41 = add(x = var_341, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_29, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_370 = const()[name = tensor("op_370"), val = tensor(0x1p-1)]; + tensor var_371 = mul(x = input_51, y = var_370)[name = tensor("op_371")]; + tensor input_53 = add(x = var_371, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_29, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_385 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([1, 5, 4, 64])]; + tensor var_387 = reshape(shape = var_386, x = var_385)[name = tensor("op_387")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_391 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_392 = const()[name = tensor("op_392"), val = tensor(0x1p-3)]; + tensor var_393 = mul(x = var_391, y = var_392)[name = tensor("op_393")]; + tensor var_394 = const()[name = tensor("op_394"), val = tensor([1, 5, 4, 64])]; + tensor var_395 = reshape(shape = var_394, x = var_393)[name = tensor("op_395")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_399 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_400 = const()[name = tensor("op_400"), val = tensor([1, 5, 4, 64])]; + tensor var_401 = reshape(shape = var_400, x = var_399)[name = tensor("op_401")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_395)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_387)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([5, 1])]; + tensor var_412 = reshape(shape = var_411, x = sqrt_s_t_3)[name = tensor("op_412")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_412)[name = tensor("M_3")]; + tensor var_414 = mul(x = qk_3, y = M_3)[name = tensor("op_414")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_401)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_414, y = v_3)[name = tensor("inner_3")]; + tensor var_416_transpose_x_0 = const()[name = tensor("op_416_transpose_x_0"), val = tensor(false)]; + tensor var_416_transpose_y_0 = const()[name = tensor("op_416_transpose_y_0"), val = tensor(false)]; + tensor var_416 = matmul(transpose_x = var_416_transpose_x_0, transpose_y = var_416_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_416")]; + tensor var_417 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_417")]; + tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1, 5, 1])]; + tensor var_419 = reshape(shape = var_418, x = var_417)[name = tensor("op_419")]; + tensor cross_3 = mul(x = var_416, y = var_419)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_422 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_422")]; + tensor var_424_transpose_x_1 = const()[name = tensor("op_424_transpose_x_1"), val = tensor(true)]; + tensor var_424_transpose_y_1 = const()[name = tensor("op_424_transpose_y_1"), val = tensor(false)]; + tensor var_424 = matmul(transpose_x = var_424_transpose_x_1, transpose_y = var_424_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_424")]; + tensor new_kv_unnorm_3 = add(x = var_422, y = var_424)[name = tensor("new_kv_unnorm_3")]; + tensor var_426 = const()[name = tensor("op_426"), val = tensor(0x1.4p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_426)[name = tensor("new_scale_3")]; + tensor var_428 = sqrt(x = new_scale_3)[name = tensor("op_428")]; + tensor var_429 = real_div(x = new_kv_unnorm_3, y = var_428)[name = tensor("op_429")]; + tensor var_430_perm_0 = const()[name = tensor("op_430_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_430 = transpose(perm = var_430_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_18, x = var_430)[name = tensor("out_9")]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 5, 256])]; + tensor out_11 = reshape(shape = var_434, x = out_9)[name = tensor("out_11")]; + tensor var_436 = silu(x = input_57)[name = tensor("op_436")]; + tensor input_59 = mul(x = var_436, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; + tensor var_444_begin_0 = const()[name = tensor("op_444_begin_0"), val = tensor([0, 0, 0])]; + tensor var_444_end_0 = const()[name = tensor("op_444_end_0"), val = tensor([1, 1, 256])]; + tensor var_444_end_mask_0 = const()[name = tensor("op_444_end_mask_0"), val = tensor([true, false, true])]; + tensor var_444 = slice_by_index(begin = var_444_begin_0, end = var_444_end_0, end_mask = var_444_end_mask_0, x = x_9)[name = tensor("op_444")]; + tensor var_447_begin_0 = const()[name = tensor("op_447_begin_0"), val = tensor([0, 1, 0])]; + tensor var_447_end_0 = const()[name = tensor("op_447_end_0"), val = tensor([1, 16, 256])]; + tensor var_447_end_mask_0 = const()[name = tensor("op_447_end_mask_0"), val = tensor([true, true, true])]; + tensor var_447 = slice_by_index(begin = var_447_begin_0, end = var_447_end_0, end_mask = var_447_end_mask_0, x = window_13)[name = tensor("op_447")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_27, interleave = window_15_interleave_0, values = (var_447, var_444))[name = tensor("window_15")]; + tensor var_452_begin_0 = const()[name = tensor("op_452_begin_0"), val = tensor([0, 1, 0])]; + tensor var_452_end_0 = const()[name = tensor("op_452_end_0"), val = tensor([1, 2, 256])]; + tensor var_452_end_mask_0 = const()[name = tensor("op_452_end_mask_0"), val = tensor([true, false, true])]; + tensor var_452 = slice_by_index(begin = var_452_begin_0, end = var_452_end_0, end_mask = var_452_end_mask_0, x = x_9)[name = tensor("op_452")]; + tensor var_455_begin_0 = const()[name = tensor("op_455_begin_0"), val = tensor([0, 1, 0])]; + tensor var_455_end_0 = const()[name = tensor("op_455_end_0"), val = tensor([1, 16, 256])]; + tensor var_455_end_mask_0 = const()[name = tensor("op_455_end_mask_0"), val = tensor([true, true, true])]; + tensor var_455 = slice_by_index(begin = var_455_begin_0, end = var_455_end_0, end_mask = var_455_end_mask_0, x = window_15)[name = tensor("op_455")]; + tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; + tensor window_17 = concat(axis = var_27, interleave = window_17_interleave_0, values = (var_455, var_452))[name = tensor("window_17")]; + tensor var_460_begin_0 = const()[name = tensor("op_460_begin_0"), val = tensor([0, 2, 0])]; + tensor var_460_end_0 = const()[name = tensor("op_460_end_0"), val = tensor([1, 3, 256])]; + tensor var_460_end_mask_0 = const()[name = tensor("op_460_end_mask_0"), val = tensor([true, false, true])]; + tensor var_460 = slice_by_index(begin = var_460_begin_0, end = var_460_end_0, end_mask = var_460_end_mask_0, x = x_9)[name = tensor("op_460")]; + tensor var_463_begin_0 = const()[name = tensor("op_463_begin_0"), val = tensor([0, 1, 0])]; + tensor var_463_end_0 = const()[name = tensor("op_463_end_0"), val = tensor([1, 16, 256])]; + tensor var_463_end_mask_0 = const()[name = tensor("op_463_end_mask_0"), val = tensor([true, true, true])]; + tensor var_463 = slice_by_index(begin = var_463_begin_0, end = var_463_end_0, end_mask = var_463_end_mask_0, x = window_17)[name = tensor("op_463")]; + tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; + tensor window_19 = concat(axis = var_27, interleave = window_19_interleave_0, values = (var_463, var_460))[name = tensor("window_19")]; + tensor var_468_begin_0 = const()[name = tensor("op_468_begin_0"), val = tensor([0, 3, 0])]; + tensor var_468_end_0 = const()[name = tensor("op_468_end_0"), val = tensor([1, 4, 256])]; + tensor var_468_end_mask_0 = const()[name = tensor("op_468_end_mask_0"), val = tensor([true, false, true])]; + tensor var_468 = slice_by_index(begin = var_468_begin_0, end = var_468_end_0, end_mask = var_468_end_mask_0, x = x_9)[name = tensor("op_468")]; + tensor var_471_begin_0 = const()[name = tensor("op_471_begin_0"), val = tensor([0, 1, 0])]; + tensor var_471_end_0 = const()[name = tensor("op_471_end_0"), val = tensor([1, 16, 256])]; + tensor var_471_end_mask_0 = const()[name = tensor("op_471_end_mask_0"), val = tensor([true, true, true])]; + tensor var_471 = slice_by_index(begin = var_471_begin_0, end = var_471_end_0, end_mask = var_471_end_mask_0, x = window_19)[name = tensor("op_471")]; + tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; + tensor window_21 = concat(axis = var_27, interleave = window_21_interleave_0, values = (var_471, var_468))[name = tensor("window_21")]; + tensor var_476_begin_0 = const()[name = tensor("op_476_begin_0"), val = tensor([0, 4, 0])]; + tensor var_476_end_0 = const()[name = tensor("op_476_end_0"), val = tensor([1, 1, 256])]; + tensor var_476_end_mask_0 = const()[name = tensor("op_476_end_mask_0"), val = tensor([true, true, true])]; + tensor var_476 = slice_by_index(begin = var_476_begin_0, end = var_476_end_0, end_mask = var_476_end_mask_0, x = x_9)[name = tensor("op_476")]; + tensor var_479_begin_0 = const()[name = tensor("op_479_begin_0"), val = tensor([0, 1, 0])]; + tensor var_479_end_0 = const()[name = tensor("op_479_end_0"), val = tensor([1, 16, 256])]; + tensor var_479_end_mask_0 = const()[name = tensor("op_479_end_mask_0"), val = tensor([true, true, true])]; + tensor var_479 = slice_by_index(begin = var_479_begin_0, end = var_479_end_0, end_mask = var_479_end_mask_0, x = window_21)[name = tensor("op_479")]; + tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; + tensor window_23 = concat(axis = var_27, interleave = window_23_interleave_0, values = (var_479, var_476))[name = tensor("window_23")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_24, interleave = input_61_interleave_0, values = (window_15, window_17, window_19, window_21, window_23))[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_504_split_sizes_0 = const()[name = tensor("op_504_split_sizes_0"), val = tensor([256, 256])]; + tensor var_504_axis_0 = const()[name = tensor("op_504_axis_0"), val = tensor(1)]; + tensor var_504_0, tensor var_504_1 = split(axis = var_504_axis_0, split_sizes = var_504_split_sizes_0, x = inputs_13)[name = tensor("op_504")]; + tensor var_506 = sigmoid(x = var_504_1)[name = tensor("op_506")]; + tensor inputs_15 = mul(x = var_504_0, y = var_506)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([5, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_537_begin_0 = const()[name = tensor("op_537_begin_0"), val = tensor([0, -1, 0])]; + tensor var_537_end_0 = const()[name = tensor("op_537_end_0"), val = tensor([5, 16, 256])]; + tensor var_537_end_mask_0 = const()[name = tensor("op_537_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_537 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = conv_out_3)[name = tensor("op_537")]; + tensor var_539_perm_0 = const()[name = tensor("op_539_perm_0"), val = tensor([1, 0, 2])]; + tensor var_539 = transpose(perm = var_539_perm_0, x = var_537)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_539)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_562 = const()[name = tensor("op_562"), val = tensor(0x1p-1)]; + tensor var_563 = mul(x = input_79, y = var_562)[name = tensor("op_563")]; + tensor input_81 = add(x = var_563, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_29, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_592 = const()[name = tensor("op_592"), val = tensor(0x1p-1)]; + tensor var_593 = mul(x = input_91, y = var_592)[name = tensor("op_593")]; + tensor input_93 = add(x = var_593, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_29, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_607 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_608 = const()[name = tensor("op_608"), val = tensor([1, 5, 4, 64])]; + tensor var_609 = reshape(shape = var_608, x = var_607)[name = tensor("op_609")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_613 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_614 = const()[name = tensor("op_614"), val = tensor(0x1p-3)]; + tensor var_615 = mul(x = var_613, y = var_614)[name = tensor("op_615")]; + tensor var_616 = const()[name = tensor("op_616"), val = tensor([1, 5, 4, 64])]; + tensor var_617 = reshape(shape = var_616, x = var_615)[name = tensor("op_617")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_621 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_622 = const()[name = tensor("op_622"), val = tensor([1, 5, 4, 64])]; + tensor var_623 = reshape(shape = var_622, x = var_621)[name = tensor("op_623")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_617)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_609)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_633 = const()[name = tensor("op_633"), val = tensor([5, 1])]; + tensor var_634 = reshape(shape = var_633, x = sqrt_s_t_5)[name = tensor("op_634")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_634)[name = tensor("M_5")]; + tensor var_636 = mul(x = qk_5, y = M_5)[name = tensor("op_636")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_623)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_636, y = v_5)[name = tensor("inner_5")]; + tensor var_638_transpose_x_0 = const()[name = tensor("op_638_transpose_x_0"), val = tensor(false)]; + tensor var_638_transpose_y_0 = const()[name = tensor("op_638_transpose_y_0"), val = tensor(false)]; + tensor var_638 = matmul(transpose_x = var_638_transpose_x_0, transpose_y = var_638_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_638")]; + tensor var_639 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_639")]; + tensor var_640 = const()[name = tensor("op_640"), val = tensor([1, 1, 5, 1])]; + tensor var_641 = reshape(shape = var_640, x = var_639)[name = tensor("op_641")]; + tensor cross_5 = mul(x = var_638, y = var_641)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_644 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_644")]; + tensor var_646_transpose_x_1 = const()[name = tensor("op_646_transpose_x_1"), val = tensor(true)]; + tensor var_646_transpose_y_1 = const()[name = tensor("op_646_transpose_y_1"), val = tensor(false)]; + tensor var_646 = matmul(transpose_x = var_646_transpose_x_1, transpose_y = var_646_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_646")]; + tensor new_kv_unnorm_5 = add(x = var_644, y = var_646)[name = tensor("new_kv_unnorm_5")]; + tensor var_648 = const()[name = tensor("op_648"), val = tensor(0x1.4p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_648)[name = tensor("new_scale_5")]; + tensor var_650 = sqrt(x = new_scale_5)[name = tensor("op_650")]; + tensor var_651 = real_div(x = new_kv_unnorm_5, y = var_650)[name = tensor("op_651")]; + tensor var_652_perm_0 = const()[name = tensor("op_652_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_652 = transpose(perm = var_652_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_18, x = var_652)[name = tensor("out_15")]; + tensor var_656 = const()[name = tensor("op_656"), val = tensor([1, 5, 256])]; + tensor out_17 = reshape(shape = var_656, x = out_15)[name = tensor("out_17")]; + tensor var_658 = silu(x = input_97)[name = tensor("op_658")]; + tensor input_99 = mul(x = var_658, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; + tensor var_666_begin_0 = const()[name = tensor("op_666_begin_0"), val = tensor([0, 0, 0])]; + tensor var_666_end_0 = const()[name = tensor("op_666_end_0"), val = tensor([1, 1, 256])]; + tensor var_666_end_mask_0 = const()[name = tensor("op_666_end_mask_0"), val = tensor([true, false, true])]; + tensor var_666 = slice_by_index(begin = var_666_begin_0, end = var_666_end_0, end_mask = var_666_end_mask_0, x = x_15)[name = tensor("op_666")]; + tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 1, 0])]; + tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 16, 256])]; + tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, true])]; + tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = window_25)[name = tensor("op_669")]; + tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; + tensor window_27 = concat(axis = var_27, interleave = window_27_interleave_0, values = (var_669, var_666))[name = tensor("window_27")]; + tensor var_674_begin_0 = const()[name = tensor("op_674_begin_0"), val = tensor([0, 1, 0])]; + tensor var_674_end_0 = const()[name = tensor("op_674_end_0"), val = tensor([1, 2, 256])]; + tensor var_674_end_mask_0 = const()[name = tensor("op_674_end_mask_0"), val = tensor([true, false, true])]; + tensor var_674 = slice_by_index(begin = var_674_begin_0, end = var_674_end_0, end_mask = var_674_end_mask_0, x = x_15)[name = tensor("op_674")]; + tensor var_677_begin_0 = const()[name = tensor("op_677_begin_0"), val = tensor([0, 1, 0])]; + tensor var_677_end_0 = const()[name = tensor("op_677_end_0"), val = tensor([1, 16, 256])]; + tensor var_677_end_mask_0 = const()[name = tensor("op_677_end_mask_0"), val = tensor([true, true, true])]; + tensor var_677 = slice_by_index(begin = var_677_begin_0, end = var_677_end_0, end_mask = var_677_end_mask_0, x = window_27)[name = tensor("op_677")]; + tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; + tensor window_29 = concat(axis = var_27, interleave = window_29_interleave_0, values = (var_677, var_674))[name = tensor("window_29")]; + tensor var_682_begin_0 = const()[name = tensor("op_682_begin_0"), val = tensor([0, 2, 0])]; + tensor var_682_end_0 = const()[name = tensor("op_682_end_0"), val = tensor([1, 3, 256])]; + tensor var_682_end_mask_0 = const()[name = tensor("op_682_end_mask_0"), val = tensor([true, false, true])]; + tensor var_682 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = x_15)[name = tensor("op_682")]; + tensor var_685_begin_0 = const()[name = tensor("op_685_begin_0"), val = tensor([0, 1, 0])]; + tensor var_685_end_0 = const()[name = tensor("op_685_end_0"), val = tensor([1, 16, 256])]; + tensor var_685_end_mask_0 = const()[name = tensor("op_685_end_mask_0"), val = tensor([true, true, true])]; + tensor var_685 = slice_by_index(begin = var_685_begin_0, end = var_685_end_0, end_mask = var_685_end_mask_0, x = window_29)[name = tensor("op_685")]; + tensor window_31_interleave_0 = const()[name = tensor("window_31_interleave_0"), val = tensor(false)]; + tensor window_31 = concat(axis = var_27, interleave = window_31_interleave_0, values = (var_685, var_682))[name = tensor("window_31")]; + tensor var_690_begin_0 = const()[name = tensor("op_690_begin_0"), val = tensor([0, 3, 0])]; + tensor var_690_end_0 = const()[name = tensor("op_690_end_0"), val = tensor([1, 4, 256])]; + tensor var_690_end_mask_0 = const()[name = tensor("op_690_end_mask_0"), val = tensor([true, false, true])]; + tensor var_690 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, x = x_15)[name = tensor("op_690")]; + tensor var_693_begin_0 = const()[name = tensor("op_693_begin_0"), val = tensor([0, 1, 0])]; + tensor var_693_end_0 = const()[name = tensor("op_693_end_0"), val = tensor([1, 16, 256])]; + tensor var_693_end_mask_0 = const()[name = tensor("op_693_end_mask_0"), val = tensor([true, true, true])]; + tensor var_693 = slice_by_index(begin = var_693_begin_0, end = var_693_end_0, end_mask = var_693_end_mask_0, x = window_31)[name = tensor("op_693")]; + tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; + tensor window_33 = concat(axis = var_27, interleave = window_33_interleave_0, values = (var_693, var_690))[name = tensor("window_33")]; + tensor var_698_begin_0 = const()[name = tensor("op_698_begin_0"), val = tensor([0, 4, 0])]; + tensor var_698_end_0 = const()[name = tensor("op_698_end_0"), val = tensor([1, 1, 256])]; + tensor var_698_end_mask_0 = const()[name = tensor("op_698_end_mask_0"), val = tensor([true, true, true])]; + tensor var_698 = slice_by_index(begin = var_698_begin_0, end = var_698_end_0, end_mask = var_698_end_mask_0, x = x_15)[name = tensor("op_698")]; + tensor var_701_begin_0 = const()[name = tensor("op_701_begin_0"), val = tensor([0, 1, 0])]; + tensor var_701_end_0 = const()[name = tensor("op_701_end_0"), val = tensor([1, 16, 256])]; + tensor var_701_end_mask_0 = const()[name = tensor("op_701_end_mask_0"), val = tensor([true, true, true])]; + tensor var_701 = slice_by_index(begin = var_701_begin_0, end = var_701_end_0, end_mask = var_701_end_mask_0, x = window_33)[name = tensor("op_701")]; + tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; + tensor window_35 = concat(axis = var_27, interleave = window_35_interleave_0, values = (var_701, var_698))[name = tensor("window_35")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_24, interleave = input_101_interleave_0, values = (window_27, window_29, window_31, window_33, window_35))[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_726_split_sizes_0 = const()[name = tensor("op_726_split_sizes_0"), val = tensor([256, 256])]; + tensor var_726_axis_0 = const()[name = tensor("op_726_axis_0"), val = tensor(1)]; + tensor var_726_0, tensor var_726_1 = split(axis = var_726_axis_0, split_sizes = var_726_split_sizes_0, x = inputs_23)[name = tensor("op_726")]; + tensor var_728 = sigmoid(x = var_726_1)[name = tensor("op_728")]; + tensor inputs_25 = mul(x = var_726_0, y = var_728)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([5, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([0, -1, 0])]; + tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([5, 16, 256])]; + tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_759 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, x = conv_out_5)[name = tensor("op_759")]; + tensor var_761_perm_0 = const()[name = tensor("op_761_perm_0"), val = tensor([1, 0, 2])]; + tensor var_761 = transpose(perm = var_761_perm_0, x = var_759)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_761)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_784 = const()[name = tensor("op_784"), val = tensor(0x1p-1)]; + tensor var_785 = mul(x = input_119, y = var_784)[name = tensor("op_785")]; + tensor input_121 = add(x = var_785, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_29, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_814 = const()[name = tensor("op_814"), val = tensor(0x1p-1)]; + tensor var_815 = mul(x = input_131, y = var_814)[name = tensor("op_815")]; + tensor input_133 = add(x = var_815, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_29, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_829 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_830 = const()[name = tensor("op_830"), val = tensor([1, 5, 4, 64])]; + tensor var_831 = reshape(shape = var_830, x = var_829)[name = tensor("op_831")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_835 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_836 = const()[name = tensor("op_836"), val = tensor(0x1p-3)]; + tensor var_837 = mul(x = var_835, y = var_836)[name = tensor("op_837")]; + tensor var_838 = const()[name = tensor("op_838"), val = tensor([1, 5, 4, 64])]; + tensor var_839 = reshape(shape = var_838, x = var_837)[name = tensor("op_839")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_843 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_844 = const()[name = tensor("op_844"), val = tensor([1, 5, 4, 64])]; + tensor var_845 = reshape(shape = var_844, x = var_843)[name = tensor("op_845")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_839)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_831)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_855 = const()[name = tensor("op_855"), val = tensor([5, 1])]; + tensor var_856 = reshape(shape = var_855, x = sqrt_s_t_7)[name = tensor("op_856")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_856)[name = tensor("M_7")]; + tensor var_858 = mul(x = qk_7, y = M_7)[name = tensor("op_858")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_845)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_858, y = v_7)[name = tensor("inner_7")]; + tensor var_860_transpose_x_0 = const()[name = tensor("op_860_transpose_x_0"), val = tensor(false)]; + tensor var_860_transpose_y_0 = const()[name = tensor("op_860_transpose_y_0"), val = tensor(false)]; + tensor var_860 = matmul(transpose_x = var_860_transpose_x_0, transpose_y = var_860_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_860")]; + tensor var_861 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_861")]; + tensor var_862 = const()[name = tensor("op_862"), val = tensor([1, 1, 5, 1])]; + tensor var_863 = reshape(shape = var_862, x = var_861)[name = tensor("op_863")]; + tensor cross_7 = mul(x = var_860, y = var_863)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_866 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_866")]; + tensor var_868_transpose_x_1 = const()[name = tensor("op_868_transpose_x_1"), val = tensor(true)]; + tensor var_868_transpose_y_1 = const()[name = tensor("op_868_transpose_y_1"), val = tensor(false)]; + tensor var_868 = matmul(transpose_x = var_868_transpose_x_1, transpose_y = var_868_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_868")]; + tensor new_kv_unnorm_7 = add(x = var_866, y = var_868)[name = tensor("new_kv_unnorm_7")]; + tensor var_870 = const()[name = tensor("op_870"), val = tensor(0x1.4p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_870)[name = tensor("new_scale_7")]; + tensor var_872 = sqrt(x = new_scale_7)[name = tensor("op_872")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_872)[name = tensor("nkv_1")]; + tensor var_874_perm_0 = const()[name = tensor("op_874_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_874 = transpose(perm = var_874_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_18, x = var_874)[name = tensor("out_21")]; + tensor var_878 = const()[name = tensor("op_878"), val = tensor([1, 5, 256])]; + tensor out_23 = reshape(shape = var_878, x = out_21)[name = tensor("out_23")]; + tensor var_880 = silu(x = input_137)[name = tensor("op_880")]; + tensor input_139 = mul(x = var_880, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_37_begin_0 = const()[name = tensor("window_37_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_37_end_0 = const()[name = tensor("window_37_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_37_end_mask_0 = const()[name = tensor("window_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_37_squeeze_mask_0 = const()[name = tensor("window_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_37 = slice_by_index(begin = window_37_begin_0, end = window_37_end_0, end_mask = window_37_end_mask_0, squeeze_mask = window_37_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_37")]; + tensor var_888_begin_0 = const()[name = tensor("op_888_begin_0"), val = tensor([0, 0, 0])]; + tensor var_888_end_0 = const()[name = tensor("op_888_end_0"), val = tensor([1, 1, 256])]; + tensor var_888_end_mask_0 = const()[name = tensor("op_888_end_mask_0"), val = tensor([true, false, true])]; + tensor var_888 = slice_by_index(begin = var_888_begin_0, end = var_888_end_0, end_mask = var_888_end_mask_0, x = x_21)[name = tensor("op_888")]; + tensor var_891_begin_0 = const()[name = tensor("op_891_begin_0"), val = tensor([0, 1, 0])]; + tensor var_891_end_0 = const()[name = tensor("op_891_end_0"), val = tensor([1, 16, 256])]; + tensor var_891_end_mask_0 = const()[name = tensor("op_891_end_mask_0"), val = tensor([true, true, true])]; + tensor var_891 = slice_by_index(begin = var_891_begin_0, end = var_891_end_0, end_mask = var_891_end_mask_0, x = window_37)[name = tensor("op_891")]; + tensor window_39_interleave_0 = const()[name = tensor("window_39_interleave_0"), val = tensor(false)]; + tensor window_39 = concat(axis = var_27, interleave = window_39_interleave_0, values = (var_891, var_888))[name = tensor("window_39")]; + tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 1, 0])]; + tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 2, 256])]; + tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, false, true])]; + tensor var_896 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = x_21)[name = tensor("op_896")]; + tensor var_899_begin_0 = const()[name = tensor("op_899_begin_0"), val = tensor([0, 1, 0])]; + tensor var_899_end_0 = const()[name = tensor("op_899_end_0"), val = tensor([1, 16, 256])]; + tensor var_899_end_mask_0 = const()[name = tensor("op_899_end_mask_0"), val = tensor([true, true, true])]; + tensor var_899 = slice_by_index(begin = var_899_begin_0, end = var_899_end_0, end_mask = var_899_end_mask_0, x = window_39)[name = tensor("op_899")]; + tensor window_41_interleave_0 = const()[name = tensor("window_41_interleave_0"), val = tensor(false)]; + tensor window_41 = concat(axis = var_27, interleave = window_41_interleave_0, values = (var_899, var_896))[name = tensor("window_41")]; + tensor var_904_begin_0 = const()[name = tensor("op_904_begin_0"), val = tensor([0, 2, 0])]; + tensor var_904_end_0 = const()[name = tensor("op_904_end_0"), val = tensor([1, 3, 256])]; + tensor var_904_end_mask_0 = const()[name = tensor("op_904_end_mask_0"), val = tensor([true, false, true])]; + tensor var_904 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = x_21)[name = tensor("op_904")]; + tensor var_907_begin_0 = const()[name = tensor("op_907_begin_0"), val = tensor([0, 1, 0])]; + tensor var_907_end_0 = const()[name = tensor("op_907_end_0"), val = tensor([1, 16, 256])]; + tensor var_907_end_mask_0 = const()[name = tensor("op_907_end_mask_0"), val = tensor([true, true, true])]; + tensor var_907 = slice_by_index(begin = var_907_begin_0, end = var_907_end_0, end_mask = var_907_end_mask_0, x = window_41)[name = tensor("op_907")]; + tensor window_43_interleave_0 = const()[name = tensor("window_43_interleave_0"), val = tensor(false)]; + tensor window_43 = concat(axis = var_27, interleave = window_43_interleave_0, values = (var_907, var_904))[name = tensor("window_43")]; + tensor var_912_begin_0 = const()[name = tensor("op_912_begin_0"), val = tensor([0, 3, 0])]; + tensor var_912_end_0 = const()[name = tensor("op_912_end_0"), val = tensor([1, 4, 256])]; + tensor var_912_end_mask_0 = const()[name = tensor("op_912_end_mask_0"), val = tensor([true, false, true])]; + tensor var_912 = slice_by_index(begin = var_912_begin_0, end = var_912_end_0, end_mask = var_912_end_mask_0, x = x_21)[name = tensor("op_912")]; + tensor var_915_begin_0 = const()[name = tensor("op_915_begin_0"), val = tensor([0, 1, 0])]; + tensor var_915_end_0 = const()[name = tensor("op_915_end_0"), val = tensor([1, 16, 256])]; + tensor var_915_end_mask_0 = const()[name = tensor("op_915_end_mask_0"), val = tensor([true, true, true])]; + tensor var_915 = slice_by_index(begin = var_915_begin_0, end = var_915_end_0, end_mask = var_915_end_mask_0, x = window_43)[name = tensor("op_915")]; + tensor window_45_interleave_0 = const()[name = tensor("window_45_interleave_0"), val = tensor(false)]; + tensor window_45 = concat(axis = var_27, interleave = window_45_interleave_0, values = (var_915, var_912))[name = tensor("window_45")]; + tensor var_920_begin_0 = const()[name = tensor("op_920_begin_0"), val = tensor([0, 4, 0])]; + tensor var_920_end_0 = const()[name = tensor("op_920_end_0"), val = tensor([1, 1, 256])]; + tensor var_920_end_mask_0 = const()[name = tensor("op_920_end_mask_0"), val = tensor([true, true, true])]; + tensor var_920 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = x_21)[name = tensor("op_920")]; + tensor var_923_begin_0 = const()[name = tensor("op_923_begin_0"), val = tensor([0, 1, 0])]; + tensor var_923_end_0 = const()[name = tensor("op_923_end_0"), val = tensor([1, 16, 256])]; + tensor var_923_end_mask_0 = const()[name = tensor("op_923_end_mask_0"), val = tensor([true, true, true])]; + tensor var_923 = slice_by_index(begin = var_923_begin_0, end = var_923_end_0, end_mask = var_923_end_mask_0, x = window_45)[name = tensor("op_923")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_27, interleave = window_interleave_0, values = (var_923, var_920))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_24, interleave = input_141_interleave_0, values = (window_39, window_41, window_43, window_45, window))[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_948_split_sizes_0 = const()[name = tensor("op_948_split_sizes_0"), val = tensor([256, 256])]; + tensor var_948_axis_0 = const()[name = tensor("op_948_axis_0"), val = tensor(1)]; + tensor var_948_0, tensor var_948_1 = split(axis = var_948_axis_0, split_sizes = var_948_split_sizes_0, x = inputs_33)[name = tensor("op_948")]; + tensor var_950 = sigmoid(x = var_948_1)[name = tensor("op_950")]; + tensor inputs_35 = mul(x = var_948_0, y = var_950)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([5, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_981_begin_0 = const()[name = tensor("op_981_begin_0"), val = tensor([0, -1, 0])]; + tensor var_981_end_0 = const()[name = tensor("op_981_end_0"), val = tensor([5, 16, 256])]; + tensor var_981_end_mask_0 = const()[name = tensor("op_981_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_981 = slice_by_index(begin = var_981_begin_0, end = var_981_end_0, end_mask = var_981_end_mask_0, x = conv_out_7)[name = tensor("op_981")]; + tensor var_983_perm_0 = const()[name = tensor("op_983_perm_0"), val = tensor([1, 0, 2])]; + tensor var_983 = transpose(perm = var_983_perm_0, x = var_981)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_983)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_1006 = const()[name = tensor("op_1006"), val = tensor(0x1p-1)]; + tensor var_1007 = mul(x = input_159, y = var_1006)[name = tensor("op_1007")]; + tensor input_161 = add(x = var_1007, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_29, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_21, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1025_begin_0 = const()[name = tensor("op_1025_begin_0"), val = tensor([0, 0, 5])]; + tensor var_1025_end_0 = const()[name = tensor("op_1025_end_0"), val = tensor([1, 256, 23])]; + tensor var_1025_end_mask_0 = const()[name = tensor("op_1025_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1025_begin_0, end = var_1025_end_0, end_mask = var_1025_end_mask_0, x = cat)[name = tensor("op_1025")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1028 = reduce_l2_norm(axes = var_1027, keep_dims = var_30, x = input_163)[name = tensor("op_1028")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_1028)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_1032_axis_0 = const()[name = tensor("op_1032_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1032_axis_0, values = (var_207, var_429, var_651, nkv_1))[name = tensor("op_1032")]; + tensor var_1034_axis_0 = const()[name = tensor("op_1034_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1034_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1034")]; + tensor var_1036_axis_0 = const()[name = tensor("op_1036_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1036_axis_0, values = (window_11, window_23, window_35, window))[name = tensor("op_1036")]; + tensor var_1045 = const()[name = tensor("op_1045"), val = tensor(0x1.5798eep-27)]; + tensor var_1050 = const()[name = tensor("op_1050"), val = tensor(0x1.4f8b58p-17)]; + tensor var_1052 = const()[name = tensor("op_1052"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_1053 = const()[name = tensor("op_1053"), val = tensor(true)]; + tensor var_1055 = const()[name = tensor("op_1055"), val = tensor(0x1p+0)]; + tensor var_1059 = const()[name = tensor("op_1059"), val = tensor(-1)]; + tensor var_1065 = const()[name = tensor("op_1065"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395712)))]; + tensor var_1127_axes_0 = const()[name = tensor("op_1127_axes_0"), val = tensor([2])]; + tensor var_1127 = expand_dims(axes = var_1127_axes_0, x = emb)[name = tensor("op_1127")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1127)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_1059, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1135_perm_0 = const()[name = tensor("op_1135_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1139 = const()[name = tensor("op_1139"), val = tensor([12, 5, 256])]; + tensor var_1135 = transpose(perm = var_1135_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1139, x = var_1135)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1147 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([12, 5, 4, 64])]; + tensor var_1149 = reshape(shape = var_1148, x = var_1147)[name = tensor("op_1149")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1153 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor(0x1p-3)]; + tensor var_1155 = mul(x = var_1153, y = var_1154)[name = tensor("op_1155")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([12, 5, 4, 64])]; + tensor var_1157 = reshape(shape = var_1156, x = var_1155)[name = tensor("op_1157")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1161 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor([12, 5, 4, 64])]; + tensor var_1163 = reshape(shape = var_1162, x = var_1161)[name = tensor("op_1163")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_1065, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_1055, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1157)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1149)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 5])]; + tensor var_1176 = reshape(shape = var_1175, x = valid_mask)[name = tensor("op_1176")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1176)[name = tensor("causal_with_valid_1")]; + tensor var_1178 = const()[name = tensor("op_1178"), val = tensor([5, 1])]; + tensor var_1179 = reshape(shape = var_1178, x = sqrt_s_t_9)[name = tensor("op_1179")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1179)[name = tensor("M_9")]; + tensor var_1181 = mul(x = qk_9, y = M_9)[name = tensor("op_1181")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1163)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1181, y = v_9)[name = tensor("inner_9")]; + tensor var_1183_transpose_x_0 = const()[name = tensor("op_1183_transpose_x_0"), val = tensor(false)]; + tensor var_1183_transpose_y_0 = const()[name = tensor("op_1183_transpose_y_0"), val = tensor(false)]; + tensor var_1183 = matmul(transpose_x = var_1183_transpose_x_0, transpose_y = var_1183_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1183")]; + tensor var_1184 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1184")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 5, 1])]; + tensor var_1186 = reshape(shape = var_1185, x = var_1184)[name = tensor("op_1186")]; + tensor cross_9 = mul(x = var_1183, y = var_1186)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1, 1, 5, 1])]; + tensor var_1190 = reshape(shape = var_1189, x = valid_mask)[name = tensor("op_1190")]; + tensor v_masked_1 = mul(x = v_9, y = var_1190)[name = tensor("v_masked_1")]; + tensor var_1192 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1192")]; + tensor var_1194_transpose_x_1 = const()[name = tensor("op_1194_transpose_x_1"), val = tensor(true)]; + tensor var_1194_transpose_y_1 = const()[name = tensor("op_1194_transpose_y_1"), val = tensor(false)]; + tensor var_1194 = matmul(transpose_x = var_1194_transpose_x_1, transpose_y = var_1194_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1194")]; + tensor new_kv_unnorm_9 = add(x = var_1192, y = var_1194)[name = tensor("new_kv_unnorm_9")]; + tensor var_1196_keep_dims_0 = const()[name = tensor("op_1196_keep_dims_0"), val = tensor(false)]; + tensor var_1196 = reduce_sum(keep_dims = var_1196_keep_dims_0, x = valid_mask)[name = tensor("op_1196")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1])]; + tensor var_1198 = reshape(shape = var_1197, x = var_1196)[name = tensor("op_1198")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1198)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_1055, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1202 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1202")]; + tensor var_1203_perm_0 = const()[name = tensor("op_1203_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1203 = transpose(perm = var_1203_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1052, x = var_1203)[name = tensor("out_27")]; + tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([12, 5, 256])]; + tensor out_29 = reshape(shape = var_1207, x = out_27)[name = tensor("out_29")]; + tensor var_1209 = silu(x = input_169)[name = tensor("op_1209")]; + tensor input_171 = mul(x = var_1209, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1050, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1219 = const()[name = tensor("op_1219"), val = tensor([1, 12, 5, 256])]; + tensor var_1220 = reshape(shape = var_1219, x = xt_1)[name = tensor("op_1220")]; + tensor var_1221_perm_0 = const()[name = tensor("op_1221_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([5, 12, 256])]; + tensor var_1221 = transpose(perm = var_1221_perm_0, x = var_1220)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1224, x = var_1221)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1247 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 5, 3, 256])]; + tensor var_1249 = reshape(shape = concat_1, x = var_1247)[name = tensor("op_1249")]; + tensor var_1250_axes_0 = const()[name = tensor("op_1250_axes_0"), val = tensor([0])]; + tensor var_1250 = expand_dims(axes = var_1250_axes_0, x = var_1249)[name = tensor("op_1250")]; + tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1252_axes_0 = const()[name = tensor("op_1252_axes_0"), val = tensor([-2])]; + tensor var_1251 = transpose(perm = var_1251_perm_0, x = var_1250)[name = tensor("transpose_21")]; + tensor var_1252 = squeeze(axes = var_1252_axes_0, x = var_1251)[name = tensor("op_1252")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 5, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1252)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 5, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1252)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 5, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1252)[name = tensor("v_11")]; + tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([12, 20, 64])]; + tensor var_1261 = reshape(shape = var_1260, x = q_11)[name = tensor("op_1261")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1267 = const()[name = tensor("op_1267"), val = tensor([12, 20, 64])]; + tensor var_1268 = reshape(shape = var_1267, x = k_11)[name = tensor("op_1268")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([12, 20, 64])]; + tensor var_1275 = reshape(shape = var_1274, x = v_11)[name = tensor("op_1275")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([5, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1261)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1278, x = q_13)[name = tensor("q_15")]; + tensor var_1280 = const()[name = tensor("op_1280"), val = tensor([5, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1268)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1280, x = k_13)[name = tensor("k_15")]; + tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([5, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1275)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1282, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1285 = const()[name = tensor("op_1285"), val = tensor([2, 0, 1, 3])]; + tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([60, 256])]; + tensor var_1286 = transpose(perm = var_1285, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1290, x = var_1286)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([12, 5, 256])]; + tensor attn_output_7 = reshape(shape = var_1294, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1050, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1050, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 5, 12, 256])]; + tensor x_31 = reshape(shape = var_1314, x = xt_3)[name = tensor("x_31")]; + tensor var_1316_perm_0 = const()[name = tensor("op_1316_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1320 = const()[name = tensor("op_1320"), val = tensor([12, 5, 256])]; + tensor var_1316 = transpose(perm = var_1316_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1320, x = var_1316)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1328 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([12, 5, 4, 64])]; + tensor var_1330 = reshape(shape = var_1329, x = var_1328)[name = tensor("op_1330")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1334 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor(0x1p-3)]; + tensor var_1336 = mul(x = var_1334, y = var_1335)[name = tensor("op_1336")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([12, 5, 4, 64])]; + tensor var_1338 = reshape(shape = var_1337, x = var_1336)[name = tensor("op_1338")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1342 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([12, 5, 4, 64])]; + tensor var_1344 = reshape(shape = var_1343, x = var_1342)[name = tensor("op_1344")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_1055, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1338)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1330)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1359 = const()[name = tensor("op_1359"), val = tensor([5, 1])]; + tensor var_1360 = reshape(shape = var_1359, x = sqrt_s_t)[name = tensor("op_1360")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1360)[name = tensor("M")]; + tensor var_1362 = mul(x = qk, y = M)[name = tensor("op_1362")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1344)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1362, y = v_17)[name = tensor("inner")]; + tensor var_1364_transpose_x_0 = const()[name = tensor("op_1364_transpose_x_0"), val = tensor(false)]; + tensor var_1364_transpose_y_0 = const()[name = tensor("op_1364_transpose_y_0"), val = tensor(false)]; + tensor var_1364 = matmul(transpose_x = var_1364_transpose_x_0, transpose_y = var_1364_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1364")]; + tensor var_1365 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1365")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 5, 1])]; + tensor var_1367 = reshape(shape = var_1366, x = var_1365)[name = tensor("op_1367")]; + tensor cross = mul(x = var_1364, y = var_1367)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1190)[name = tensor("v_masked")]; + tensor var_1373 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1373")]; + tensor var_1375_transpose_x_1 = const()[name = tensor("op_1375_transpose_x_1"), val = tensor(true)]; + tensor var_1375_transpose_y_1 = const()[name = tensor("op_1375_transpose_y_1"), val = tensor(false)]; + tensor var_1375 = matmul(transpose_x = var_1375_transpose_x_1, transpose_y = var_1375_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1375")]; + tensor new_kv_unnorm = add(x = var_1373, y = var_1375)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1198)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_1055, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1384_perm_0 = const()[name = tensor("op_1384_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1384 = transpose(perm = var_1384_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1052, x = var_1384)[name = tensor("out_33")]; + tensor var_1388 = const()[name = tensor("op_1388"), val = tensor([12, 5, 256])]; + tensor out = reshape(shape = var_1388, x = out_33)[name = tensor("out")]; + tensor var_1390 = silu(x = input_187)[name = tensor("op_1390")]; + tensor input_189 = mul(x = var_1390, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1050, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 12, 5, 256])]; + tensor var_1401 = reshape(shape = var_1400, x = xt_5)[name = tensor("op_1401")]; + tensor var_1402_perm_0 = const()[name = tensor("op_1402_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1405 = const()[name = tensor("op_1405"), val = tensor([5, 12, 256])]; + tensor var_1402 = transpose(perm = var_1402_perm_0, x = var_1401)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1405, x = var_1402)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1428 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 5, 3, 256])]; + tensor var_1430 = reshape(shape = concat_2, x = var_1428)[name = tensor("op_1430")]; + tensor var_1431_axes_0 = const()[name = tensor("op_1431_axes_0"), val = tensor([0])]; + tensor var_1431 = expand_dims(axes = var_1431_axes_0, x = var_1430)[name = tensor("op_1431")]; + tensor var_1432_perm_0 = const()[name = tensor("op_1432_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1433_axes_0 = const()[name = tensor("op_1433_axes_0"), val = tensor([-2])]; + tensor var_1432 = transpose(perm = var_1432_perm_0, x = var_1431)[name = tensor("transpose_8")]; + tensor var_1433 = squeeze(axes = var_1433_axes_0, x = var_1432)[name = tensor("op_1433")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 5, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1433)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 5, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1433)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 5, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1433)[name = tensor("v_19")]; + tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([12, 20, 64])]; + tensor var_1442 = reshape(shape = var_1441, x = q_19)[name = tensor("op_1442")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([12, 20, 64])]; + tensor var_1449 = reshape(shape = var_1448, x = k_19)[name = tensor("op_1449")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([12, 20, 64])]; + tensor var_1456 = reshape(shape = var_1455, x = v_19)[name = tensor("op_1456")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([5, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1442)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1459, x = q_21)[name = tensor("q")]; + tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([5, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1449)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1461, x = k_21)[name = tensor("k")]; + tensor var_1463 = const()[name = tensor("op_1463"), val = tensor([5, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1456)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1463, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1466 = const()[name = tensor("op_1466"), val = tensor([2, 0, 1, 3])]; + tensor var_1471 = const()[name = tensor("op_1471"), val = tensor([60, 256])]; + tensor var_1467 = transpose(perm = var_1466, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1471, x = var_1467)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1475 = const()[name = tensor("op_1475"), val = tensor([12, 5, 256])]; + tensor attn_output = reshape(shape = var_1475, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1050, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1050, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1495 = const()[name = tensor("op_1495"), val = tensor([1, 5, 12, 256])]; + tensor input = reshape(shape = var_1495, x = xt)[name = tensor("input")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([-1])]; + tensor var_1498 = reduce_l2_norm(axes = var_1497, keep_dims = var_1053, x = input)[name = tensor("op_1498")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_1045, beta = const_42, x = var_1498)[name = tensor("clip_5")]; + tensor var_1500 = real_div(x = input, y = clip_5)[name = tensor("op_1500")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([5, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([5, 256, 12])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1500)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 5, 12])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 5, 11])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1504")]; + tensor var_1506_axis_0 = const()[name = tensor("op_1506_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1506_axis_0, values = (var_1202, nkv))[name = tensor("op_1506")]; + tensor var_1508_axis_0 = const()[name = tensor("op_1508_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1508_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1508")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/weights/weight.bin b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..3b4f535f87fa314558508dec7a3666c51d985b8f --- /dev/null +++ b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6cf5ed3f25665bacb776dbed74c04cb1a932ecafe85e66b1222af05baed824a +size 44457216 diff --git a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..ee2f5685e43d7adb38c9c1a4e1b73168d29549bc --- /dev/null +++ b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c600d793bc84760911328536382b994a486396bf5f73c044403146818f2ba2f1 +size 196627 diff --git a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..3b4f535f87fa314558508dec7a3666c51d985b8f --- /dev/null +++ b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6cf5ed3f25665bacb776dbed74c04cb1a932ecafe85e66b1222af05baed824a +size 44457216 diff --git a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Manifest.json b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..9b05f7f5a11617791e1ce810a13372eb75f0afe2 --- /dev/null +++ b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "4275B100-4E00-4F6A-9426-B138B856397B": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "DBD9F036-EA3D-4076-9E45-DF3A8F67CEED": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "4275B100-4E00-4F6A-9426-B138B856397B" +} diff --git a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/analytics/coremldata.bin b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..009646e23619a556a10590ffca6efa31fb7f6eeb --- /dev/null +++ b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23998be748af73a6dec0a0dced0634a6a32bf25edbeb9b6f24938033cbb5bcbf +size 243 diff --git a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/coremldata.bin b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c569cfb25eae428b97226bb2c29697a5083561fb --- /dev/null +++ b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:868aa7ddffd7f41be3a3dc214b15d2e85acbbceb685c24d452f5c4629170b995 +size 1310 diff --git a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/metadata.json b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..74e418759f83a521fd3e5c06595da4efc184dc64 --- /dev/null +++ b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND DIHARD III streaming diarizer (pipeline, T=1, max_speakers=10)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 10)", + "shortDescription" : "", + "shape" : "[1, 1, 10]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 66, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 36, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 10, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 46 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 345)", + "shortDescription" : "", + "shape" : "[1, 1, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"dih3\", \"model_label\": \"DIHARD III\", \"variant\": \"pipeline\", \"chunk_size\": 1, \"step_duration_ms\": 100, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_dih3_100ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/model.mil b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..182f88767b6a7156fba8d5e17cc71543704b1b0a --- /dev/null +++ b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/model.mil @@ -0,0 +1,1184 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0]])]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 1, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 1, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 1, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 1, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+0)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 1, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, x_3))[name = tensor("window_3")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = window_3)[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_249_split_sizes_0 = const()[name = tensor("op_249_split_sizes_0"), val = tensor([256, 256])]; + tensor var_249_axis_0 = const()[name = tensor("op_249_axis_0"), val = tensor(1)]; + tensor var_249_0, tensor var_249_1 = split(axis = var_249_axis_0, split_sizes = var_249_split_sizes_0, x = inputs_3)[name = tensor("op_249")]; + tensor var_251 = sigmoid(x = var_249_1)[name = tensor("op_251")]; + tensor inputs_5 = mul(x = var_249_0, y = var_251)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([1, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_282_begin_0 = const()[name = tensor("op_282_begin_0"), val = tensor([0, -1, 0])]; + tensor var_282_end_0 = const()[name = tensor("op_282_end_0"), val = tensor([1, 16, 256])]; + tensor var_282_end_mask_0 = const()[name = tensor("op_282_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_282 = slice_by_index(begin = var_282_begin_0, end = var_282_end_0, end_mask = var_282_end_mask_0, x = conv_out_1)[name = tensor("op_282")]; + tensor var_284_perm_0 = const()[name = tensor("op_284_perm_0"), val = tensor([1, 0, 2])]; + tensor var_284 = transpose(perm = var_284_perm_0, x = var_282)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_284)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_307 = const()[name = tensor("op_307"), val = tensor(0x1p-1)]; + tensor var_308 = mul(x = input_39, y = var_307)[name = tensor("op_308")]; + tensor input_41 = add(x = var_308, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_337 = const()[name = tensor("op_337"), val = tensor(0x1p-1)]; + tensor var_338 = mul(x = input_51, y = var_337)[name = tensor("op_338")]; + tensor input_53 = add(x = var_338, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_352 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_353 = const()[name = tensor("op_353"), val = tensor([1, 1, 4, 64])]; + tensor var_354 = reshape(shape = var_353, x = var_352)[name = tensor("op_354")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_358 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_359 = const()[name = tensor("op_359"), val = tensor(0x1p-3)]; + tensor var_360 = mul(x = var_358, y = var_359)[name = tensor("op_360")]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 1, 4, 64])]; + tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_366 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_367 = const()[name = tensor("op_367"), val = tensor([1, 1, 4, 64])]; + tensor var_368 = reshape(shape = var_367, x = var_366)[name = tensor("op_368")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_362)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_354)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 1])]; + tensor var_379 = reshape(shape = var_378, x = sqrt_s_t_3)[name = tensor("op_379")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_379)[name = tensor("M_3")]; + tensor var_381 = mul(x = qk_3, y = M_3)[name = tensor("op_381")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_368)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_381, y = v_3)[name = tensor("inner_3")]; + tensor var_383_transpose_x_0 = const()[name = tensor("op_383_transpose_x_0"), val = tensor(false)]; + tensor var_383_transpose_y_0 = const()[name = tensor("op_383_transpose_y_0"), val = tensor(false)]; + tensor var_383 = matmul(transpose_x = var_383_transpose_x_0, transpose_y = var_383_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_383")]; + tensor var_384 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_384")]; + tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 1, 1, 1])]; + tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; + tensor cross_3 = mul(x = var_383, y = var_386)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_389 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_389")]; + tensor var_391_transpose_x_1 = const()[name = tensor("op_391_transpose_x_1"), val = tensor(true)]; + tensor var_391_transpose_y_1 = const()[name = tensor("op_391_transpose_y_1"), val = tensor(false)]; + tensor var_391 = matmul(transpose_x = var_391_transpose_x_1, transpose_y = var_391_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_391")]; + tensor new_kv_unnorm_3 = add(x = var_389, y = var_391)[name = tensor("new_kv_unnorm_3")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor(0x1p+0)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_393)[name = tensor("new_scale_3")]; + tensor var_395 = sqrt(x = new_scale_3)[name = tensor("op_395")]; + tensor var_396 = real_div(x = new_kv_unnorm_3, y = var_395)[name = tensor("op_396")]; + tensor var_397_perm_0 = const()[name = tensor("op_397_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_397 = transpose(perm = var_397_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_397)[name = tensor("out_9")]; + tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 256])]; + tensor out_11 = reshape(shape = var_401, x = out_9)[name = tensor("out_11")]; + tensor var_403 = silu(x = input_57)[name = tensor("op_403")]; + tensor input_59 = mul(x = var_403, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_5_begin_0 = const()[name = tensor("window_5_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_5_end_0 = const()[name = tensor("window_5_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_5_end_mask_0 = const()[name = tensor("window_5_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_5_squeeze_mask_0 = const()[name = tensor("window_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_5 = slice_by_index(begin = window_5_begin_0, end = window_5_end_0, end_mask = window_5_end_mask_0, squeeze_mask = window_5_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_5")]; + tensor var_414_begin_0 = const()[name = tensor("op_414_begin_0"), val = tensor([0, 1, 0])]; + tensor var_414_end_0 = const()[name = tensor("op_414_end_0"), val = tensor([1, 16, 256])]; + tensor var_414_end_mask_0 = const()[name = tensor("op_414_end_mask_0"), val = tensor([true, true, true])]; + tensor var_414 = slice_by_index(begin = var_414_begin_0, end = var_414_end_0, end_mask = var_414_end_mask_0, x = window_5)[name = tensor("op_414")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_414, x_9))[name = tensor("window_7")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = window_7)[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_439_split_sizes_0 = const()[name = tensor("op_439_split_sizes_0"), val = tensor([256, 256])]; + tensor var_439_axis_0 = const()[name = tensor("op_439_axis_0"), val = tensor(1)]; + tensor var_439_0, tensor var_439_1 = split(axis = var_439_axis_0, split_sizes = var_439_split_sizes_0, x = inputs_13)[name = tensor("op_439")]; + tensor var_441 = sigmoid(x = var_439_1)[name = tensor("op_441")]; + tensor inputs_15 = mul(x = var_439_0, y = var_441)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([1, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_472_begin_0 = const()[name = tensor("op_472_begin_0"), val = tensor([0, -1, 0])]; + tensor var_472_end_0 = const()[name = tensor("op_472_end_0"), val = tensor([1, 16, 256])]; + tensor var_472_end_mask_0 = const()[name = tensor("op_472_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_472 = slice_by_index(begin = var_472_begin_0, end = var_472_end_0, end_mask = var_472_end_mask_0, x = conv_out_3)[name = tensor("op_472")]; + tensor var_474_perm_0 = const()[name = tensor("op_474_perm_0"), val = tensor([1, 0, 2])]; + tensor var_474 = transpose(perm = var_474_perm_0, x = var_472)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_474)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_497 = const()[name = tensor("op_497"), val = tensor(0x1p-1)]; + tensor var_498 = mul(x = input_79, y = var_497)[name = tensor("op_498")]; + tensor input_81 = add(x = var_498, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_527 = const()[name = tensor("op_527"), val = tensor(0x1p-1)]; + tensor var_528 = mul(x = input_91, y = var_527)[name = tensor("op_528")]; + tensor input_93 = add(x = var_528, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_542 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 1, 4, 64])]; + tensor var_544 = reshape(shape = var_543, x = var_542)[name = tensor("op_544")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_548 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_549 = const()[name = tensor("op_549"), val = tensor(0x1p-3)]; + tensor var_550 = mul(x = var_548, y = var_549)[name = tensor("op_550")]; + tensor var_551 = const()[name = tensor("op_551"), val = tensor([1, 1, 4, 64])]; + tensor var_552 = reshape(shape = var_551, x = var_550)[name = tensor("op_552")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_556 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_557 = const()[name = tensor("op_557"), val = tensor([1, 1, 4, 64])]; + tensor var_558 = reshape(shape = var_557, x = var_556)[name = tensor("op_558")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_552)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_544)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_568 = const()[name = tensor("op_568"), val = tensor([1, 1])]; + tensor var_569 = reshape(shape = var_568, x = sqrt_s_t_5)[name = tensor("op_569")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_569)[name = tensor("M_5")]; + tensor var_571 = mul(x = qk_5, y = M_5)[name = tensor("op_571")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_558)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_571, y = v_5)[name = tensor("inner_5")]; + tensor var_573_transpose_x_0 = const()[name = tensor("op_573_transpose_x_0"), val = tensor(false)]; + tensor var_573_transpose_y_0 = const()[name = tensor("op_573_transpose_y_0"), val = tensor(false)]; + tensor var_573 = matmul(transpose_x = var_573_transpose_x_0, transpose_y = var_573_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_573")]; + tensor var_574 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_574")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 1, 1, 1])]; + tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; + tensor cross_5 = mul(x = var_573, y = var_576)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_579 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_579")]; + tensor var_581_transpose_x_1 = const()[name = tensor("op_581_transpose_x_1"), val = tensor(true)]; + tensor var_581_transpose_y_1 = const()[name = tensor("op_581_transpose_y_1"), val = tensor(false)]; + tensor var_581 = matmul(transpose_x = var_581_transpose_x_1, transpose_y = var_581_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_581")]; + tensor new_kv_unnorm_5 = add(x = var_579, y = var_581)[name = tensor("new_kv_unnorm_5")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor(0x1p+0)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_583)[name = tensor("new_scale_5")]; + tensor var_585 = sqrt(x = new_scale_5)[name = tensor("op_585")]; + tensor var_586 = real_div(x = new_kv_unnorm_5, y = var_585)[name = tensor("op_586")]; + tensor var_587_perm_0 = const()[name = tensor("op_587_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_587 = transpose(perm = var_587_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_587)[name = tensor("out_15")]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 256])]; + tensor out_17 = reshape(shape = var_591, x = out_15)[name = tensor("out_17")]; + tensor var_593 = silu(x = input_97)[name = tensor("op_593")]; + tensor input_99 = mul(x = var_593, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; + tensor var_604_begin_0 = const()[name = tensor("op_604_begin_0"), val = tensor([0, 1, 0])]; + tensor var_604_end_0 = const()[name = tensor("op_604_end_0"), val = tensor([1, 16, 256])]; + tensor var_604_end_mask_0 = const()[name = tensor("op_604_end_mask_0"), val = tensor([true, true, true])]; + tensor var_604 = slice_by_index(begin = var_604_begin_0, end = var_604_end_0, end_mask = var_604_end_mask_0, x = window_9)[name = tensor("op_604")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_604, x_15))[name = tensor("window_11")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = window_11)[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_629_split_sizes_0 = const()[name = tensor("op_629_split_sizes_0"), val = tensor([256, 256])]; + tensor var_629_axis_0 = const()[name = tensor("op_629_axis_0"), val = tensor(1)]; + tensor var_629_0, tensor var_629_1 = split(axis = var_629_axis_0, split_sizes = var_629_split_sizes_0, x = inputs_23)[name = tensor("op_629")]; + tensor var_631 = sigmoid(x = var_629_1)[name = tensor("op_631")]; + tensor inputs_25 = mul(x = var_629_0, y = var_631)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([1, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_662_begin_0 = const()[name = tensor("op_662_begin_0"), val = tensor([0, -1, 0])]; + tensor var_662_end_0 = const()[name = tensor("op_662_end_0"), val = tensor([1, 16, 256])]; + tensor var_662_end_mask_0 = const()[name = tensor("op_662_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_662 = slice_by_index(begin = var_662_begin_0, end = var_662_end_0, end_mask = var_662_end_mask_0, x = conv_out_5)[name = tensor("op_662")]; + tensor var_664_perm_0 = const()[name = tensor("op_664_perm_0"), val = tensor([1, 0, 2])]; + tensor var_664 = transpose(perm = var_664_perm_0, x = var_662)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_664)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_687 = const()[name = tensor("op_687"), val = tensor(0x1p-1)]; + tensor var_688 = mul(x = input_119, y = var_687)[name = tensor("op_688")]; + tensor input_121 = add(x = var_688, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_717 = const()[name = tensor("op_717"), val = tensor(0x1p-1)]; + tensor var_718 = mul(x = input_131, y = var_717)[name = tensor("op_718")]; + tensor input_133 = add(x = var_718, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_732 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_733 = const()[name = tensor("op_733"), val = tensor([1, 1, 4, 64])]; + tensor var_734 = reshape(shape = var_733, x = var_732)[name = tensor("op_734")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_738 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_739 = const()[name = tensor("op_739"), val = tensor(0x1p-3)]; + tensor var_740 = mul(x = var_738, y = var_739)[name = tensor("op_740")]; + tensor var_741 = const()[name = tensor("op_741"), val = tensor([1, 1, 4, 64])]; + tensor var_742 = reshape(shape = var_741, x = var_740)[name = tensor("op_742")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_746 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 1, 4, 64])]; + tensor var_748 = reshape(shape = var_747, x = var_746)[name = tensor("op_748")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_742)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_734)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_758 = const()[name = tensor("op_758"), val = tensor([1, 1])]; + tensor var_759 = reshape(shape = var_758, x = sqrt_s_t_7)[name = tensor("op_759")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_759)[name = tensor("M_7")]; + tensor var_761 = mul(x = qk_7, y = M_7)[name = tensor("op_761")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_748)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_761, y = v_7)[name = tensor("inner_7")]; + tensor var_763_transpose_x_0 = const()[name = tensor("op_763_transpose_x_0"), val = tensor(false)]; + tensor var_763_transpose_y_0 = const()[name = tensor("op_763_transpose_y_0"), val = tensor(false)]; + tensor var_763 = matmul(transpose_x = var_763_transpose_x_0, transpose_y = var_763_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_763")]; + tensor var_764 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_764")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 1, 1, 1])]; + tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; + tensor cross_7 = mul(x = var_763, y = var_766)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_769 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_769")]; + tensor var_771_transpose_x_1 = const()[name = tensor("op_771_transpose_x_1"), val = tensor(true)]; + tensor var_771_transpose_y_1 = const()[name = tensor("op_771_transpose_y_1"), val = tensor(false)]; + tensor var_771 = matmul(transpose_x = var_771_transpose_x_1, transpose_y = var_771_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_771")]; + tensor new_kv_unnorm_7 = add(x = var_769, y = var_771)[name = tensor("new_kv_unnorm_7")]; + tensor var_773 = const()[name = tensor("op_773"), val = tensor(0x1p+0)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_773)[name = tensor("new_scale_7")]; + tensor var_775 = sqrt(x = new_scale_7)[name = tensor("op_775")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_775)[name = tensor("nkv_1")]; + tensor var_777_perm_0 = const()[name = tensor("op_777_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_777 = transpose(perm = var_777_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_777)[name = tensor("out_21")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 1, 256])]; + tensor out_23 = reshape(shape = var_781, x = out_21)[name = tensor("out_23")]; + tensor var_783 = silu(x = input_137)[name = tensor("op_783")]; + tensor input_139 = mul(x = var_783, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; + tensor var_794_begin_0 = const()[name = tensor("op_794_begin_0"), val = tensor([0, 1, 0])]; + tensor var_794_end_0 = const()[name = tensor("op_794_end_0"), val = tensor([1, 16, 256])]; + tensor var_794_end_mask_0 = const()[name = tensor("op_794_end_mask_0"), val = tensor([true, true, true])]; + tensor var_794 = slice_by_index(begin = var_794_begin_0, end = var_794_end_0, end_mask = var_794_end_mask_0, x = window_13)[name = tensor("op_794")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_794, x_21))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = window)[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_819_split_sizes_0 = const()[name = tensor("op_819_split_sizes_0"), val = tensor([256, 256])]; + tensor var_819_axis_0 = const()[name = tensor("op_819_axis_0"), val = tensor(1)]; + tensor var_819_0, tensor var_819_1 = split(axis = var_819_axis_0, split_sizes = var_819_split_sizes_0, x = inputs_33)[name = tensor("op_819")]; + tensor var_821 = sigmoid(x = var_819_1)[name = tensor("op_821")]; + tensor inputs_35 = mul(x = var_819_0, y = var_821)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([1, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_852_begin_0 = const()[name = tensor("op_852_begin_0"), val = tensor([0, -1, 0])]; + tensor var_852_end_0 = const()[name = tensor("op_852_end_0"), val = tensor([1, 16, 256])]; + tensor var_852_end_mask_0 = const()[name = tensor("op_852_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_852 = slice_by_index(begin = var_852_begin_0, end = var_852_end_0, end_mask = var_852_end_mask_0, x = conv_out_7)[name = tensor("op_852")]; + tensor var_854_perm_0 = const()[name = tensor("op_854_perm_0"), val = tensor([1, 0, 2])]; + tensor var_854 = transpose(perm = var_854_perm_0, x = var_852)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_854)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_877 = const()[name = tensor("op_877"), val = tensor(0x1p-1)]; + tensor var_878 = mul(x = input_159, y = var_877)[name = tensor("op_878")]; + tensor input_161 = add(x = var_878, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 0, 1])]; + tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 256, 19])]; + tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = cat)[name = tensor("op_896")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_898 = const()[name = tensor("op_898"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_899 = reduce_l2_norm(axes = var_898, keep_dims = var_29, x = input_163)[name = tensor("op_899")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_899)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_903_axis_0 = const()[name = tensor("op_903_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_903_axis_0, values = (var_206, var_396, var_586, nkv_1))[name = tensor("op_903")]; + tensor var_905_axis_0 = const()[name = tensor("op_905_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_905_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_905")]; + tensor var_907_axis_0 = const()[name = tensor("op_907_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_907_axis_0, values = (window_3, window_7, window_11, window))[name = tensor("op_907")]; + tensor var_916 = const()[name = tensor("op_916"), val = tensor(0x1.5798eep-27)]; + tensor var_921 = const()[name = tensor("op_921"), val = tensor(0x1.4f8b58p-17)]; + tensor var_923 = const()[name = tensor("op_923"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_924 = const()[name = tensor("op_924"), val = tensor(true)]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor(0x1p+0)]; + tensor var_930 = const()[name = tensor("op_930"), val = tensor(-1)]; + tensor var_936 = const()[name = tensor("op_936"), val = tensor(0)]; + tensor var_993 = const()[name = tensor("op_993"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_998_axes_0 = const()[name = tensor("op_998_axes_0"), val = tensor([2])]; + tensor var_998 = expand_dims(axes = var_998_axes_0, x = emb)[name = tensor("op_998")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_998)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_930, interleave = input_165_interleave_0, values = (emb_exp, var_993))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1006_perm_0 = const()[name = tensor("op_1006_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([12, 1, 256])]; + tensor var_1006 = transpose(perm = var_1006_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1010, x = var_1006)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1018 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1019 = const()[name = tensor("op_1019"), val = tensor([12, 1, 4, 64])]; + tensor var_1020 = reshape(shape = var_1019, x = var_1018)[name = tensor("op_1020")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1024 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1025 = const()[name = tensor("op_1025"), val = tensor(0x1p-3)]; + tensor var_1026 = mul(x = var_1024, y = var_1025)[name = tensor("op_1026")]; + tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([12, 1, 4, 64])]; + tensor var_1028 = reshape(shape = var_1027, x = var_1026)[name = tensor("op_1028")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1032 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1033 = const()[name = tensor("op_1033"), val = tensor([12, 1, 4, 64])]; + tensor var_1034 = reshape(shape = var_1033, x = var_1032)[name = tensor("op_1034")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_936, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_926, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1028)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1020)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 1])]; + tensor var_1047 = reshape(shape = var_1046, x = valid_mask)[name = tensor("op_1047")]; + tensor var_1049 = const()[name = tensor("op_1049"), val = tensor([1, 1])]; + tensor var_1050 = reshape(shape = var_1049, x = sqrt_s_t_9)[name = tensor("op_1050")]; + tensor M_9 = real_div(x = var_1047, y = var_1050)[name = tensor("M_9")]; + tensor var_1052 = mul(x = qk_9, y = M_9)[name = tensor("op_1052")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1034)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1052, y = v_9)[name = tensor("inner_9")]; + tensor var_1054_transpose_x_0 = const()[name = tensor("op_1054_transpose_x_0"), val = tensor(false)]; + tensor var_1054_transpose_y_0 = const()[name = tensor("op_1054_transpose_y_0"), val = tensor(false)]; + tensor var_1054 = matmul(transpose_x = var_1054_transpose_x_0, transpose_y = var_1054_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1054")]; + tensor var_1055 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1055")]; + tensor var_1056 = const()[name = tensor("op_1056"), val = tensor([1, 1, 1, 1])]; + tensor var_1057 = reshape(shape = var_1056, x = var_1055)[name = tensor("op_1057")]; + tensor cross_9 = mul(x = var_1054, y = var_1057)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1060 = const()[name = tensor("op_1060"), val = tensor([1, 1, 1, 1])]; + tensor var_1061 = reshape(shape = var_1060, x = valid_mask)[name = tensor("op_1061")]; + tensor v_masked_1 = mul(x = v_9, y = var_1061)[name = tensor("v_masked_1")]; + tensor var_1063 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1063")]; + tensor var_1065_transpose_x_1 = const()[name = tensor("op_1065_transpose_x_1"), val = tensor(true)]; + tensor var_1065_transpose_y_1 = const()[name = tensor("op_1065_transpose_y_1"), val = tensor(false)]; + tensor var_1065 = matmul(transpose_x = var_1065_transpose_x_1, transpose_y = var_1065_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1065")]; + tensor new_kv_unnorm_9 = add(x = var_1063, y = var_1065)[name = tensor("new_kv_unnorm_9")]; + tensor var_1067_keep_dims_0 = const()[name = tensor("op_1067_keep_dims_0"), val = tensor(false)]; + tensor var_1067 = reduce_sum(keep_dims = var_1067_keep_dims_0, x = valid_mask)[name = tensor("op_1067")]; + tensor var_1068 = const()[name = tensor("op_1068"), val = tensor([1])]; + tensor var_1069 = reshape(shape = var_1068, x = var_1067)[name = tensor("op_1069")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1069)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_926, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1073 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1073")]; + tensor var_1074_perm_0 = const()[name = tensor("op_1074_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1074 = transpose(perm = var_1074_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_923, x = var_1074)[name = tensor("out_27")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([12, 1, 256])]; + tensor out_29 = reshape(shape = var_1078, x = out_27)[name = tensor("out_29")]; + tensor var_1080 = silu(x = input_169)[name = tensor("op_1080")]; + tensor input_171 = mul(x = var_1080, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_921, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1090 = const()[name = tensor("op_1090"), val = tensor([1, 12, 1, 256])]; + tensor var_1091 = reshape(shape = var_1090, x = xt_1)[name = tensor("op_1091")]; + tensor var_1092_perm_0 = const()[name = tensor("op_1092_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1095 = const()[name = tensor("op_1095"), val = tensor([1, 12, 256])]; + tensor var_1092 = transpose(perm = var_1092_perm_0, x = var_1091)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1095, x = var_1092)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1118 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 1, 3, 256])]; + tensor var_1120 = reshape(shape = concat_1, x = var_1118)[name = tensor("op_1120")]; + tensor var_1121_axes_0 = const()[name = tensor("op_1121_axes_0"), val = tensor([0])]; + tensor var_1121 = expand_dims(axes = var_1121_axes_0, x = var_1120)[name = tensor("op_1121")]; + tensor var_1122_perm_0 = const()[name = tensor("op_1122_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1123_axes_0 = const()[name = tensor("op_1123_axes_0"), val = tensor([-2])]; + tensor var_1122 = transpose(perm = var_1122_perm_0, x = var_1121)[name = tensor("transpose_21")]; + tensor var_1123 = squeeze(axes = var_1123_axes_0, x = var_1122)[name = tensor("op_1123")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 1, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1123)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 1, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1123)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 1, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1123)[name = tensor("v_11")]; + tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([12, 4, 64])]; + tensor var_1132 = reshape(shape = var_1131, x = q_11)[name = tensor("op_1132")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1138 = const()[name = tensor("op_1138"), val = tensor([12, 4, 64])]; + tensor var_1139 = reshape(shape = var_1138, x = k_11)[name = tensor("op_1139")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([12, 4, 64])]; + tensor var_1146 = reshape(shape = var_1145, x = v_11)[name = tensor("op_1146")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1149 = const()[name = tensor("op_1149"), val = tensor([1, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1132)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1149, x = q_13)[name = tensor("q_15")]; + tensor var_1151 = const()[name = tensor("op_1151"), val = tensor([1, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1139)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1151, x = k_13)[name = tensor("k_15")]; + tensor var_1153 = const()[name = tensor("op_1153"), val = tensor([1, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1146)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1153, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([2, 0, 1, 3])]; + tensor var_1161 = const()[name = tensor("op_1161"), val = tensor([12, 256])]; + tensor var_1157 = transpose(perm = var_1156, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1161, x = var_1157)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1165 = const()[name = tensor("op_1165"), val = tensor([12, 1, 256])]; + tensor attn_output_7 = reshape(shape = var_1165, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_921, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_921, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 12, 256])]; + tensor x_31 = reshape(shape = var_1185, x = xt_3)[name = tensor("x_31")]; + tensor var_1187_perm_0 = const()[name = tensor("op_1187_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([12, 1, 256])]; + tensor var_1187 = transpose(perm = var_1187_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1191, x = var_1187)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1199 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1200 = const()[name = tensor("op_1200"), val = tensor([12, 1, 4, 64])]; + tensor var_1201 = reshape(shape = var_1200, x = var_1199)[name = tensor("op_1201")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1205 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1206 = const()[name = tensor("op_1206"), val = tensor(0x1p-3)]; + tensor var_1207 = mul(x = var_1205, y = var_1206)[name = tensor("op_1207")]; + tensor var_1208 = const()[name = tensor("op_1208"), val = tensor([12, 1, 4, 64])]; + tensor var_1209 = reshape(shape = var_1208, x = var_1207)[name = tensor("op_1209")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1213 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1214 = const()[name = tensor("op_1214"), val = tensor([12, 1, 4, 64])]; + tensor var_1215 = reshape(shape = var_1214, x = var_1213)[name = tensor("op_1215")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_926, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1209)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1201)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 1])]; + tensor var_1231 = reshape(shape = var_1230, x = sqrt_s_t)[name = tensor("op_1231")]; + tensor M = real_div(x = var_1047, y = var_1231)[name = tensor("M")]; + tensor var_1233 = mul(x = qk, y = M)[name = tensor("op_1233")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1215)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1233, y = v_17)[name = tensor("inner")]; + tensor var_1235_transpose_x_0 = const()[name = tensor("op_1235_transpose_x_0"), val = tensor(false)]; + tensor var_1235_transpose_y_0 = const()[name = tensor("op_1235_transpose_y_0"), val = tensor(false)]; + tensor var_1235 = matmul(transpose_x = var_1235_transpose_x_0, transpose_y = var_1235_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1235")]; + tensor var_1236 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1236")]; + tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1, 1, 1])]; + tensor var_1238 = reshape(shape = var_1237, x = var_1236)[name = tensor("op_1238")]; + tensor cross = mul(x = var_1235, y = var_1238)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1061)[name = tensor("v_masked")]; + tensor var_1244 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1244")]; + tensor var_1246_transpose_x_1 = const()[name = tensor("op_1246_transpose_x_1"), val = tensor(true)]; + tensor var_1246_transpose_y_1 = const()[name = tensor("op_1246_transpose_y_1"), val = tensor(false)]; + tensor var_1246 = matmul(transpose_x = var_1246_transpose_x_1, transpose_y = var_1246_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1246")]; + tensor new_kv_unnorm = add(x = var_1244, y = var_1246)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1069)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_926, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1255_perm_0 = const()[name = tensor("op_1255_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1255 = transpose(perm = var_1255_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_923, x = var_1255)[name = tensor("out_33")]; + tensor var_1259 = const()[name = tensor("op_1259"), val = tensor([12, 1, 256])]; + tensor out = reshape(shape = var_1259, x = out_33)[name = tensor("out")]; + tensor var_1261 = silu(x = input_187)[name = tensor("op_1261")]; + tensor input_189 = mul(x = var_1261, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_921, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1271 = const()[name = tensor("op_1271"), val = tensor([1, 12, 1, 256])]; + tensor var_1272 = reshape(shape = var_1271, x = xt_5)[name = tensor("op_1272")]; + tensor var_1273_perm_0 = const()[name = tensor("op_1273_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1276 = const()[name = tensor("op_1276"), val = tensor([1, 12, 256])]; + tensor var_1273 = transpose(perm = var_1273_perm_0, x = var_1272)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1276, x = var_1273)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1299 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 1, 3, 256])]; + tensor var_1301 = reshape(shape = concat_2, x = var_1299)[name = tensor("op_1301")]; + tensor var_1302_axes_0 = const()[name = tensor("op_1302_axes_0"), val = tensor([0])]; + tensor var_1302 = expand_dims(axes = var_1302_axes_0, x = var_1301)[name = tensor("op_1302")]; + tensor var_1303_perm_0 = const()[name = tensor("op_1303_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1304_axes_0 = const()[name = tensor("op_1304_axes_0"), val = tensor([-2])]; + tensor var_1303 = transpose(perm = var_1303_perm_0, x = var_1302)[name = tensor("transpose_8")]; + tensor var_1304 = squeeze(axes = var_1304_axes_0, x = var_1303)[name = tensor("op_1304")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 1, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1304)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 1, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1304)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 1, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1304)[name = tensor("v_19")]; + tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([12, 4, 64])]; + tensor var_1313 = reshape(shape = var_1312, x = q_19)[name = tensor("op_1313")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1319 = const()[name = tensor("op_1319"), val = tensor([12, 4, 64])]; + tensor var_1320 = reshape(shape = var_1319, x = k_19)[name = tensor("op_1320")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([12, 4, 64])]; + tensor var_1327 = reshape(shape = var_1326, x = v_19)[name = tensor("op_1327")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1330 = const()[name = tensor("op_1330"), val = tensor([1, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1313)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1330, x = q_21)[name = tensor("q")]; + tensor var_1332 = const()[name = tensor("op_1332"), val = tensor([1, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1320)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1332, x = k_21)[name = tensor("k")]; + tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1327)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1334, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([2, 0, 1, 3])]; + tensor var_1342 = const()[name = tensor("op_1342"), val = tensor([12, 256])]; + tensor var_1338 = transpose(perm = var_1337, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1342, x = var_1338)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1346 = const()[name = tensor("op_1346"), val = tensor([12, 1, 256])]; + tensor attn_output = reshape(shape = var_1346, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_921, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_921, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 12, 256])]; + tensor input = reshape(shape = var_1366, x = xt)[name = tensor("input")]; + tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([-1])]; + tensor var_1369 = reduce_l2_norm(axes = var_1368, keep_dims = var_924, x = input)[name = tensor("op_1369")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_916, beta = const_42, x = var_1369)[name = tensor("clip_5")]; + tensor var_1371 = real_div(x = input, y = clip_5)[name = tensor("op_1371")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([1, 256, 12])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1371)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = emb, y = reshape_1)[name = tensor("matmul_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 1, 11])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = matmul_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1375")]; + tensor var_1377_axis_0 = const()[name = tensor("op_1377_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1377_axis_0, values = (var_1073, nkv))[name = tensor("op_1377")]; + tensor var_1379_axis_0 = const()[name = tensor("op_1379_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1379_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1379")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/weights/weight.bin b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..025a1557bf7fbe59d082831d46abc7b5b9d6cd08 --- /dev/null +++ b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bf5dc575c0dcb1856eb21e25f401e9e45c183c663f6b6324404b650fdcdbec2 +size 44407680 diff --git a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..172bd36e29739117a908278b3525a67e992d8b17 --- /dev/null +++ b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:866e7d0226752a637b1394b3ae9ecf1dc1de0ca1073f1b164eb38c8dfe62ea4a +size 171366 diff --git a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..025a1557bf7fbe59d082831d46abc7b5b9d6cd08 --- /dev/null +++ b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bf5dc575c0dcb1856eb21e25f401e9e45c183c663f6b6324404b650fdcdbec2 +size 44407680 diff --git a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Manifest.json b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..5e96fabcd00f0d81e3d7405b594d3c65e54a58a8 --- /dev/null +++ b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "80A6D93A-77B4-40EA-A15C-5545A78FEEBD": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "9DDE2145-36B0-46F7-8DFC-385831DC4F71": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "80A6D93A-77B4-40EA-A15C-5545A78FEEBD" +} diff --git a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/analytics/coremldata.bin b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..5409bd7627c8339408f2ba44c89764b14943a6db --- /dev/null +++ b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8893bce28227e8d51b3751b9be3a44e3ec395d16b2515ff8db197af2749801d +size 243 diff --git a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/coremldata.bin b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..1ab9910e40d04ee5bb87f2f64a0f79e96a63d69c --- /dev/null +++ b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:714cac00c71f584a18c944c841c3236d25c7bded9052f77aa3c110e1c0373194 +size 1310 diff --git a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/metadata.json b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..aaab8ad1bf6a40dc63cb5e669c82cca3eee7f301 --- /dev/null +++ b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND DIHARD III streaming diarizer (pipeline, T=2, max_speakers=10)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 2 × 10)", + "shortDescription" : "", + "shape" : "[1, 2, 10]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 68, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 48, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 14, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 47 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 2 × 345)", + "shortDescription" : "", + "shape" : "[1, 2, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2)", + "shortDescription" : "", + "shape" : "[2]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"dih3\", \"model_label\": \"DIHARD III\", \"variant\": \"pipeline\", \"chunk_size\": 2, \"step_duration_ms\": 200, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_dih3_200ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/model.mil b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..5651eaf71d9adaeffd8f5bfe34de4f9950314977 --- /dev/null +++ b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/model.mil @@ -0,0 +1,1246 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 2, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 2, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 2, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([2, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 2, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 2, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; + tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; + tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; + tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; + tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; + tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 1, 256])]; + tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, true, true])]; + tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; + tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; + tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; + tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; + tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5))[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_257_split_sizes_0 = const()[name = tensor("op_257_split_sizes_0"), val = tensor([256, 256])]; + tensor var_257_axis_0 = const()[name = tensor("op_257_axis_0"), val = tensor(1)]; + tensor var_257_0, tensor var_257_1 = split(axis = var_257_axis_0, split_sizes = var_257_split_sizes_0, x = inputs_3)[name = tensor("op_257")]; + tensor var_259 = sigmoid(x = var_257_1)[name = tensor("op_259")]; + tensor inputs_5 = mul(x = var_257_0, y = var_259)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([2, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_290_begin_0 = const()[name = tensor("op_290_begin_0"), val = tensor([0, -1, 0])]; + tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([2, 16, 256])]; + tensor var_290_end_mask_0 = const()[name = tensor("op_290_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_290 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, x = conv_out_1)[name = tensor("op_290")]; + tensor var_292_perm_0 = const()[name = tensor("op_292_perm_0"), val = tensor([1, 0, 2])]; + tensor var_292 = transpose(perm = var_292_perm_0, x = var_290)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_292)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_315 = const()[name = tensor("op_315"), val = tensor(0x1p-1)]; + tensor var_316 = mul(x = input_39, y = var_315)[name = tensor("op_316")]; + tensor input_41 = add(x = var_316, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_345 = const()[name = tensor("op_345"), val = tensor(0x1p-1)]; + tensor var_346 = mul(x = input_51, y = var_345)[name = tensor("op_346")]; + tensor input_53 = add(x = var_346, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_360 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 2, 4, 64])]; + tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_366 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_367 = const()[name = tensor("op_367"), val = tensor(0x1p-3)]; + tensor var_368 = mul(x = var_366, y = var_367)[name = tensor("op_368")]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 2, 4, 64])]; + tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_374 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_375 = const()[name = tensor("op_375"), val = tensor([1, 2, 4, 64])]; + tensor var_376 = reshape(shape = var_375, x = var_374)[name = tensor("op_376")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_370)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_362)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([2, 1])]; + tensor var_387 = reshape(shape = var_386, x = sqrt_s_t_3)[name = tensor("op_387")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_387)[name = tensor("M_3")]; + tensor var_389 = mul(x = qk_3, y = M_3)[name = tensor("op_389")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_376)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_389, y = v_3)[name = tensor("inner_3")]; + tensor var_391_transpose_x_0 = const()[name = tensor("op_391_transpose_x_0"), val = tensor(false)]; + tensor var_391_transpose_y_0 = const()[name = tensor("op_391_transpose_y_0"), val = tensor(false)]; + tensor var_391 = matmul(transpose_x = var_391_transpose_x_0, transpose_y = var_391_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_391")]; + tensor var_392 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_392")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor([1, 1, 2, 1])]; + tensor var_394 = reshape(shape = var_393, x = var_392)[name = tensor("op_394")]; + tensor cross_3 = mul(x = var_391, y = var_394)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_397 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_397")]; + tensor var_399_transpose_x_1 = const()[name = tensor("op_399_transpose_x_1"), val = tensor(true)]; + tensor var_399_transpose_y_1 = const()[name = tensor("op_399_transpose_y_1"), val = tensor(false)]; + tensor var_399 = matmul(transpose_x = var_399_transpose_x_1, transpose_y = var_399_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_399")]; + tensor new_kv_unnorm_3 = add(x = var_397, y = var_399)[name = tensor("new_kv_unnorm_3")]; + tensor var_401 = const()[name = tensor("op_401"), val = tensor(0x1p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_401)[name = tensor("new_scale_3")]; + tensor var_403 = sqrt(x = new_scale_3)[name = tensor("op_403")]; + tensor var_404 = real_div(x = new_kv_unnorm_3, y = var_403)[name = tensor("op_404")]; + tensor var_405_perm_0 = const()[name = tensor("op_405_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_405 = transpose(perm = var_405_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_405)[name = tensor("out_9")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 2, 256])]; + tensor out_11 = reshape(shape = var_409, x = out_9)[name = tensor("out_11")]; + tensor var_411 = silu(x = input_57)[name = tensor("op_411")]; + tensor input_59 = mul(x = var_411, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_7_begin_0 = const()[name = tensor("window_7_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_7_end_0 = const()[name = tensor("window_7_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_7_end_mask_0 = const()[name = tensor("window_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_7_squeeze_mask_0 = const()[name = tensor("window_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_7 = slice_by_index(begin = window_7_begin_0, end = window_7_end_0, end_mask = window_7_end_mask_0, squeeze_mask = window_7_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_7")]; + tensor var_419_begin_0 = const()[name = tensor("op_419_begin_0"), val = tensor([0, 0, 0])]; + tensor var_419_end_0 = const()[name = tensor("op_419_end_0"), val = tensor([1, 1, 256])]; + tensor var_419_end_mask_0 = const()[name = tensor("op_419_end_mask_0"), val = tensor([true, false, true])]; + tensor var_419 = slice_by_index(begin = var_419_begin_0, end = var_419_end_0, end_mask = var_419_end_mask_0, x = x_9)[name = tensor("op_419")]; + tensor var_422_begin_0 = const()[name = tensor("op_422_begin_0"), val = tensor([0, 1, 0])]; + tensor var_422_end_0 = const()[name = tensor("op_422_end_0"), val = tensor([1, 16, 256])]; + tensor var_422_end_mask_0 = const()[name = tensor("op_422_end_mask_0"), val = tensor([true, true, true])]; + tensor var_422 = slice_by_index(begin = var_422_begin_0, end = var_422_end_0, end_mask = var_422_end_mask_0, x = window_7)[name = tensor("op_422")]; + tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; + tensor window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_422, var_419))[name = tensor("window_9")]; + tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 1, 0])]; + tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; + tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, true, true])]; + tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; + tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; + tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; + tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; + tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_9, window_11))[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_455_split_sizes_0 = const()[name = tensor("op_455_split_sizes_0"), val = tensor([256, 256])]; + tensor var_455_axis_0 = const()[name = tensor("op_455_axis_0"), val = tensor(1)]; + tensor var_455_0, tensor var_455_1 = split(axis = var_455_axis_0, split_sizes = var_455_split_sizes_0, x = inputs_13)[name = tensor("op_455")]; + tensor var_457 = sigmoid(x = var_455_1)[name = tensor("op_457")]; + tensor inputs_15 = mul(x = var_455_0, y = var_457)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([2, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_488_begin_0 = const()[name = tensor("op_488_begin_0"), val = tensor([0, -1, 0])]; + tensor var_488_end_0 = const()[name = tensor("op_488_end_0"), val = tensor([2, 16, 256])]; + tensor var_488_end_mask_0 = const()[name = tensor("op_488_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_488 = slice_by_index(begin = var_488_begin_0, end = var_488_end_0, end_mask = var_488_end_mask_0, x = conv_out_3)[name = tensor("op_488")]; + tensor var_490_perm_0 = const()[name = tensor("op_490_perm_0"), val = tensor([1, 0, 2])]; + tensor var_490 = transpose(perm = var_490_perm_0, x = var_488)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_490)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_513 = const()[name = tensor("op_513"), val = tensor(0x1p-1)]; + tensor var_514 = mul(x = input_79, y = var_513)[name = tensor("op_514")]; + tensor input_81 = add(x = var_514, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_543 = const()[name = tensor("op_543"), val = tensor(0x1p-1)]; + tensor var_544 = mul(x = input_91, y = var_543)[name = tensor("op_544")]; + tensor input_93 = add(x = var_544, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_558 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_559 = const()[name = tensor("op_559"), val = tensor([1, 2, 4, 64])]; + tensor var_560 = reshape(shape = var_559, x = var_558)[name = tensor("op_560")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_564 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_565 = const()[name = tensor("op_565"), val = tensor(0x1p-3)]; + tensor var_566 = mul(x = var_564, y = var_565)[name = tensor("op_566")]; + tensor var_567 = const()[name = tensor("op_567"), val = tensor([1, 2, 4, 64])]; + tensor var_568 = reshape(shape = var_567, x = var_566)[name = tensor("op_568")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_572 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 2, 4, 64])]; + tensor var_574 = reshape(shape = var_573, x = var_572)[name = tensor("op_574")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_568)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_560)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_584 = const()[name = tensor("op_584"), val = tensor([2, 1])]; + tensor var_585 = reshape(shape = var_584, x = sqrt_s_t_5)[name = tensor("op_585")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_585)[name = tensor("M_5")]; + tensor var_587 = mul(x = qk_5, y = M_5)[name = tensor("op_587")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_574)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_587, y = v_5)[name = tensor("inner_5")]; + tensor var_589_transpose_x_0 = const()[name = tensor("op_589_transpose_x_0"), val = tensor(false)]; + tensor var_589_transpose_y_0 = const()[name = tensor("op_589_transpose_y_0"), val = tensor(false)]; + tensor var_589 = matmul(transpose_x = var_589_transpose_x_0, transpose_y = var_589_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_589")]; + tensor var_590 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_590")]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 2, 1])]; + tensor var_592 = reshape(shape = var_591, x = var_590)[name = tensor("op_592")]; + tensor cross_5 = mul(x = var_589, y = var_592)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_595 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_595")]; + tensor var_597_transpose_x_1 = const()[name = tensor("op_597_transpose_x_1"), val = tensor(true)]; + tensor var_597_transpose_y_1 = const()[name = tensor("op_597_transpose_y_1"), val = tensor(false)]; + tensor var_597 = matmul(transpose_x = var_597_transpose_x_1, transpose_y = var_597_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_597")]; + tensor new_kv_unnorm_5 = add(x = var_595, y = var_597)[name = tensor("new_kv_unnorm_5")]; + tensor var_599 = const()[name = tensor("op_599"), val = tensor(0x1p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_599)[name = tensor("new_scale_5")]; + tensor var_601 = sqrt(x = new_scale_5)[name = tensor("op_601")]; + tensor var_602 = real_div(x = new_kv_unnorm_5, y = var_601)[name = tensor("op_602")]; + tensor var_603_perm_0 = const()[name = tensor("op_603_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_603 = transpose(perm = var_603_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_603)[name = tensor("out_15")]; + tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 2, 256])]; + tensor out_17 = reshape(shape = var_607, x = out_15)[name = tensor("out_17")]; + tensor var_609 = silu(x = input_97)[name = tensor("op_609")]; + tensor input_99 = mul(x = var_609, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; + tensor var_617_begin_0 = const()[name = tensor("op_617_begin_0"), val = tensor([0, 0, 0])]; + tensor var_617_end_0 = const()[name = tensor("op_617_end_0"), val = tensor([1, 1, 256])]; + tensor var_617_end_mask_0 = const()[name = tensor("op_617_end_mask_0"), val = tensor([true, false, true])]; + tensor var_617 = slice_by_index(begin = var_617_begin_0, end = var_617_end_0, end_mask = var_617_end_mask_0, x = x_15)[name = tensor("op_617")]; + tensor var_620_begin_0 = const()[name = tensor("op_620_begin_0"), val = tensor([0, 1, 0])]; + tensor var_620_end_0 = const()[name = tensor("op_620_end_0"), val = tensor([1, 16, 256])]; + tensor var_620_end_mask_0 = const()[name = tensor("op_620_end_mask_0"), val = tensor([true, true, true])]; + tensor var_620 = slice_by_index(begin = var_620_begin_0, end = var_620_end_0, end_mask = var_620_end_mask_0, x = window_13)[name = tensor("op_620")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_620, var_617))[name = tensor("window_15")]; + tensor var_625_begin_0 = const()[name = tensor("op_625_begin_0"), val = tensor([0, 1, 0])]; + tensor var_625_end_0 = const()[name = tensor("op_625_end_0"), val = tensor([1, 1, 256])]; + tensor var_625_end_mask_0 = const()[name = tensor("op_625_end_mask_0"), val = tensor([true, true, true])]; + tensor var_625 = slice_by_index(begin = var_625_begin_0, end = var_625_end_0, end_mask = var_625_end_mask_0, x = x_15)[name = tensor("op_625")]; + tensor var_628_begin_0 = const()[name = tensor("op_628_begin_0"), val = tensor([0, 1, 0])]; + tensor var_628_end_0 = const()[name = tensor("op_628_end_0"), val = tensor([1, 16, 256])]; + tensor var_628_end_mask_0 = const()[name = tensor("op_628_end_mask_0"), val = tensor([true, true, true])]; + tensor var_628 = slice_by_index(begin = var_628_begin_0, end = var_628_end_0, end_mask = var_628_end_mask_0, x = window_15)[name = tensor("op_628")]; + tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; + tensor window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_628, var_625))[name = tensor("window_17")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_15, window_17))[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_653_split_sizes_0 = const()[name = tensor("op_653_split_sizes_0"), val = tensor([256, 256])]; + tensor var_653_axis_0 = const()[name = tensor("op_653_axis_0"), val = tensor(1)]; + tensor var_653_0, tensor var_653_1 = split(axis = var_653_axis_0, split_sizes = var_653_split_sizes_0, x = inputs_23)[name = tensor("op_653")]; + tensor var_655 = sigmoid(x = var_653_1)[name = tensor("op_655")]; + tensor inputs_25 = mul(x = var_653_0, y = var_655)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([2, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_686_begin_0 = const()[name = tensor("op_686_begin_0"), val = tensor([0, -1, 0])]; + tensor var_686_end_0 = const()[name = tensor("op_686_end_0"), val = tensor([2, 16, 256])]; + tensor var_686_end_mask_0 = const()[name = tensor("op_686_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_686 = slice_by_index(begin = var_686_begin_0, end = var_686_end_0, end_mask = var_686_end_mask_0, x = conv_out_5)[name = tensor("op_686")]; + tensor var_688_perm_0 = const()[name = tensor("op_688_perm_0"), val = tensor([1, 0, 2])]; + tensor var_688 = transpose(perm = var_688_perm_0, x = var_686)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_688)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_711 = const()[name = tensor("op_711"), val = tensor(0x1p-1)]; + tensor var_712 = mul(x = input_119, y = var_711)[name = tensor("op_712")]; + tensor input_121 = add(x = var_712, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_741 = const()[name = tensor("op_741"), val = tensor(0x1p-1)]; + tensor var_742 = mul(x = input_131, y = var_741)[name = tensor("op_742")]; + tensor input_133 = add(x = var_742, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_756 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_757 = const()[name = tensor("op_757"), val = tensor([1, 2, 4, 64])]; + tensor var_758 = reshape(shape = var_757, x = var_756)[name = tensor("op_758")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_762 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_763 = const()[name = tensor("op_763"), val = tensor(0x1p-3)]; + tensor var_764 = mul(x = var_762, y = var_763)[name = tensor("op_764")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 2, 4, 64])]; + tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_770 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 2, 4, 64])]; + tensor var_772 = reshape(shape = var_771, x = var_770)[name = tensor("op_772")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_766)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_758)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_782 = const()[name = tensor("op_782"), val = tensor([2, 1])]; + tensor var_783 = reshape(shape = var_782, x = sqrt_s_t_7)[name = tensor("op_783")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_783)[name = tensor("M_7")]; + tensor var_785 = mul(x = qk_7, y = M_7)[name = tensor("op_785")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_772)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_785, y = v_7)[name = tensor("inner_7")]; + tensor var_787_transpose_x_0 = const()[name = tensor("op_787_transpose_x_0"), val = tensor(false)]; + tensor var_787_transpose_y_0 = const()[name = tensor("op_787_transpose_y_0"), val = tensor(false)]; + tensor var_787 = matmul(transpose_x = var_787_transpose_x_0, transpose_y = var_787_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_787")]; + tensor var_788 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_788")]; + tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 1, 2, 1])]; + tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; + tensor cross_7 = mul(x = var_787, y = var_790)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_793 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_793")]; + tensor var_795_transpose_x_1 = const()[name = tensor("op_795_transpose_x_1"), val = tensor(true)]; + tensor var_795_transpose_y_1 = const()[name = tensor("op_795_transpose_y_1"), val = tensor(false)]; + tensor var_795 = matmul(transpose_x = var_795_transpose_x_1, transpose_y = var_795_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_795")]; + tensor new_kv_unnorm_7 = add(x = var_793, y = var_795)[name = tensor("new_kv_unnorm_7")]; + tensor var_797 = const()[name = tensor("op_797"), val = tensor(0x1p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_797)[name = tensor("new_scale_7")]; + tensor var_799 = sqrt(x = new_scale_7)[name = tensor("op_799")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_799)[name = tensor("nkv_1")]; + tensor var_801_perm_0 = const()[name = tensor("op_801_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_801 = transpose(perm = var_801_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_801)[name = tensor("out_21")]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 2, 256])]; + tensor out_23 = reshape(shape = var_805, x = out_21)[name = tensor("out_23")]; + tensor var_807 = silu(x = input_137)[name = tensor("op_807")]; + tensor input_139 = mul(x = var_807, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_19_begin_0 = const()[name = tensor("window_19_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_19_end_0 = const()[name = tensor("window_19_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_19_end_mask_0 = const()[name = tensor("window_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_19_squeeze_mask_0 = const()[name = tensor("window_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_19 = slice_by_index(begin = window_19_begin_0, end = window_19_end_0, end_mask = window_19_end_mask_0, squeeze_mask = window_19_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_19")]; + tensor var_815_begin_0 = const()[name = tensor("op_815_begin_0"), val = tensor([0, 0, 0])]; + tensor var_815_end_0 = const()[name = tensor("op_815_end_0"), val = tensor([1, 1, 256])]; + tensor var_815_end_mask_0 = const()[name = tensor("op_815_end_mask_0"), val = tensor([true, false, true])]; + tensor var_815 = slice_by_index(begin = var_815_begin_0, end = var_815_end_0, end_mask = var_815_end_mask_0, x = x_21)[name = tensor("op_815")]; + tensor var_818_begin_0 = const()[name = tensor("op_818_begin_0"), val = tensor([0, 1, 0])]; + tensor var_818_end_0 = const()[name = tensor("op_818_end_0"), val = tensor([1, 16, 256])]; + tensor var_818_end_mask_0 = const()[name = tensor("op_818_end_mask_0"), val = tensor([true, true, true])]; + tensor var_818 = slice_by_index(begin = var_818_begin_0, end = var_818_end_0, end_mask = var_818_end_mask_0, x = window_19)[name = tensor("op_818")]; + tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; + tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_818, var_815))[name = tensor("window_21")]; + tensor var_823_begin_0 = const()[name = tensor("op_823_begin_0"), val = tensor([0, 1, 0])]; + tensor var_823_end_0 = const()[name = tensor("op_823_end_0"), val = tensor([1, 1, 256])]; + tensor var_823_end_mask_0 = const()[name = tensor("op_823_end_mask_0"), val = tensor([true, true, true])]; + tensor var_823 = slice_by_index(begin = var_823_begin_0, end = var_823_end_0, end_mask = var_823_end_mask_0, x = x_21)[name = tensor("op_823")]; + tensor var_826_begin_0 = const()[name = tensor("op_826_begin_0"), val = tensor([0, 1, 0])]; + tensor var_826_end_0 = const()[name = tensor("op_826_end_0"), val = tensor([1, 16, 256])]; + tensor var_826_end_mask_0 = const()[name = tensor("op_826_end_mask_0"), val = tensor([true, true, true])]; + tensor var_826 = slice_by_index(begin = var_826_begin_0, end = var_826_end_0, end_mask = var_826_end_mask_0, x = window_21)[name = tensor("op_826")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_826, var_823))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_21, window))[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_851_split_sizes_0 = const()[name = tensor("op_851_split_sizes_0"), val = tensor([256, 256])]; + tensor var_851_axis_0 = const()[name = tensor("op_851_axis_0"), val = tensor(1)]; + tensor var_851_0, tensor var_851_1 = split(axis = var_851_axis_0, split_sizes = var_851_split_sizes_0, x = inputs_33)[name = tensor("op_851")]; + tensor var_853 = sigmoid(x = var_851_1)[name = tensor("op_853")]; + tensor inputs_35 = mul(x = var_851_0, y = var_853)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([2, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_884_begin_0 = const()[name = tensor("op_884_begin_0"), val = tensor([0, -1, 0])]; + tensor var_884_end_0 = const()[name = tensor("op_884_end_0"), val = tensor([2, 16, 256])]; + tensor var_884_end_mask_0 = const()[name = tensor("op_884_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_884 = slice_by_index(begin = var_884_begin_0, end = var_884_end_0, end_mask = var_884_end_mask_0, x = conv_out_7)[name = tensor("op_884")]; + tensor var_886_perm_0 = const()[name = tensor("op_886_perm_0"), val = tensor([1, 0, 2])]; + tensor var_886 = transpose(perm = var_886_perm_0, x = var_884)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_886)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_909 = const()[name = tensor("op_909"), val = tensor(0x1p-1)]; + tensor var_910 = mul(x = input_159, y = var_909)[name = tensor("op_910")]; + tensor input_161 = add(x = var_910, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_928_begin_0 = const()[name = tensor("op_928_begin_0"), val = tensor([0, 0, 2])]; + tensor var_928_end_0 = const()[name = tensor("op_928_end_0"), val = tensor([1, 256, 20])]; + tensor var_928_end_mask_0 = const()[name = tensor("op_928_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_928_begin_0, end = var_928_end_0, end_mask = var_928_end_mask_0, x = cat)[name = tensor("op_928")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_930 = const()[name = tensor("op_930"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_931 = reduce_l2_norm(axes = var_930, keep_dims = var_29, x = input_163)[name = tensor("op_931")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_931)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_935_axis_0 = const()[name = tensor("op_935_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_935_axis_0, values = (var_206, var_404, var_602, nkv_1))[name = tensor("op_935")]; + tensor var_937_axis_0 = const()[name = tensor("op_937_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_937_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_937")]; + tensor var_939_axis_0 = const()[name = tensor("op_939_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_939_axis_0, values = (window_5, window_11, window_17, window))[name = tensor("op_939")]; + tensor var_948 = const()[name = tensor("op_948"), val = tensor(0x1.5798eep-27)]; + tensor var_953 = const()[name = tensor("op_953"), val = tensor(0x1.4f8b58p-17)]; + tensor var_955 = const()[name = tensor("op_955"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_956 = const()[name = tensor("op_956"), val = tensor(true)]; + tensor var_958 = const()[name = tensor("op_958"), val = tensor(0x1p+0)]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor(-1)]; + tensor var_968 = const()[name = tensor("op_968"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_1030_axes_0 = const()[name = tensor("op_1030_axes_0"), val = tensor([2])]; + tensor var_1030 = expand_dims(axes = var_1030_axes_0, x = emb)[name = tensor("op_1030")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1030)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_962, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1038_perm_0 = const()[name = tensor("op_1038_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1042 = const()[name = tensor("op_1042"), val = tensor([12, 2, 256])]; + tensor var_1038 = transpose(perm = var_1038_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1042, x = var_1038)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1050 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1051 = const()[name = tensor("op_1051"), val = tensor([12, 2, 4, 64])]; + tensor var_1052 = reshape(shape = var_1051, x = var_1050)[name = tensor("op_1052")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1056 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1057 = const()[name = tensor("op_1057"), val = tensor(0x1p-3)]; + tensor var_1058 = mul(x = var_1056, y = var_1057)[name = tensor("op_1058")]; + tensor var_1059 = const()[name = tensor("op_1059"), val = tensor([12, 2, 4, 64])]; + tensor var_1060 = reshape(shape = var_1059, x = var_1058)[name = tensor("op_1060")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1064 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1065 = const()[name = tensor("op_1065"), val = tensor([12, 2, 4, 64])]; + tensor var_1066 = reshape(shape = var_1065, x = var_1064)[name = tensor("op_1066")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_968, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_958, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1060)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1052)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([1, 2])]; + tensor var_1079 = reshape(shape = var_1078, x = valid_mask)[name = tensor("op_1079")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1079)[name = tensor("causal_with_valid_1")]; + tensor var_1081 = const()[name = tensor("op_1081"), val = tensor([2, 1])]; + tensor var_1082 = reshape(shape = var_1081, x = sqrt_s_t_9)[name = tensor("op_1082")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1082)[name = tensor("M_9")]; + tensor var_1084 = mul(x = qk_9, y = M_9)[name = tensor("op_1084")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1066)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1084, y = v_9)[name = tensor("inner_9")]; + tensor var_1086_transpose_x_0 = const()[name = tensor("op_1086_transpose_x_0"), val = tensor(false)]; + tensor var_1086_transpose_y_0 = const()[name = tensor("op_1086_transpose_y_0"), val = tensor(false)]; + tensor var_1086 = matmul(transpose_x = var_1086_transpose_x_0, transpose_y = var_1086_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1086")]; + tensor var_1087 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1087")]; + tensor var_1088 = const()[name = tensor("op_1088"), val = tensor([1, 1, 2, 1])]; + tensor var_1089 = reshape(shape = var_1088, x = var_1087)[name = tensor("op_1089")]; + tensor cross_9 = mul(x = var_1086, y = var_1089)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([1, 1, 2, 1])]; + tensor var_1093 = reshape(shape = var_1092, x = valid_mask)[name = tensor("op_1093")]; + tensor v_masked_1 = mul(x = v_9, y = var_1093)[name = tensor("v_masked_1")]; + tensor var_1095 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1095")]; + tensor var_1097_transpose_x_1 = const()[name = tensor("op_1097_transpose_x_1"), val = tensor(true)]; + tensor var_1097_transpose_y_1 = const()[name = tensor("op_1097_transpose_y_1"), val = tensor(false)]; + tensor var_1097 = matmul(transpose_x = var_1097_transpose_x_1, transpose_y = var_1097_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1097")]; + tensor new_kv_unnorm_9 = add(x = var_1095, y = var_1097)[name = tensor("new_kv_unnorm_9")]; + tensor var_1099_keep_dims_0 = const()[name = tensor("op_1099_keep_dims_0"), val = tensor(false)]; + tensor var_1099 = reduce_sum(keep_dims = var_1099_keep_dims_0, x = valid_mask)[name = tensor("op_1099")]; + tensor var_1100 = const()[name = tensor("op_1100"), val = tensor([1])]; + tensor var_1101 = reshape(shape = var_1100, x = var_1099)[name = tensor("op_1101")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1101)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_958, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1105 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1105")]; + tensor var_1106_perm_0 = const()[name = tensor("op_1106_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1106 = transpose(perm = var_1106_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_955, x = var_1106)[name = tensor("out_27")]; + tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([12, 2, 256])]; + tensor out_29 = reshape(shape = var_1110, x = out_27)[name = tensor("out_29")]; + tensor var_1112 = silu(x = input_169)[name = tensor("op_1112")]; + tensor input_171 = mul(x = var_1112, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_953, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1122 = const()[name = tensor("op_1122"), val = tensor([1, 12, 2, 256])]; + tensor var_1123 = reshape(shape = var_1122, x = xt_1)[name = tensor("op_1123")]; + tensor var_1124_perm_0 = const()[name = tensor("op_1124_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([2, 12, 256])]; + tensor var_1124 = transpose(perm = var_1124_perm_0, x = var_1123)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1127, x = var_1124)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1150 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 2, 3, 256])]; + tensor var_1152 = reshape(shape = concat_1, x = var_1150)[name = tensor("op_1152")]; + tensor var_1153_axes_0 = const()[name = tensor("op_1153_axes_0"), val = tensor([0])]; + tensor var_1153 = expand_dims(axes = var_1153_axes_0, x = var_1152)[name = tensor("op_1153")]; + tensor var_1154_perm_0 = const()[name = tensor("op_1154_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1155_axes_0 = const()[name = tensor("op_1155_axes_0"), val = tensor([-2])]; + tensor var_1154 = transpose(perm = var_1154_perm_0, x = var_1153)[name = tensor("transpose_21")]; + tensor var_1155 = squeeze(axes = var_1155_axes_0, x = var_1154)[name = tensor("op_1155")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 2, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1155)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 2, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1155)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 2, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1155)[name = tensor("v_11")]; + tensor var_1163 = const()[name = tensor("op_1163"), val = tensor([12, 8, 64])]; + tensor var_1164 = reshape(shape = var_1163, x = q_11)[name = tensor("op_1164")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1170 = const()[name = tensor("op_1170"), val = tensor([12, 8, 64])]; + tensor var_1171 = reshape(shape = var_1170, x = k_11)[name = tensor("op_1171")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor([12, 8, 64])]; + tensor var_1178 = reshape(shape = var_1177, x = v_11)[name = tensor("op_1178")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1181 = const()[name = tensor("op_1181"), val = tensor([2, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1164)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1181, x = q_13)[name = tensor("q_15")]; + tensor var_1183 = const()[name = tensor("op_1183"), val = tensor([2, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1171)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1183, x = k_13)[name = tensor("k_15")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([2, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1178)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1185, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1188 = const()[name = tensor("op_1188"), val = tensor([2, 0, 1, 3])]; + tensor var_1193 = const()[name = tensor("op_1193"), val = tensor([24, 256])]; + tensor var_1189 = transpose(perm = var_1188, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1193, x = var_1189)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([12, 2, 256])]; + tensor attn_output_7 = reshape(shape = var_1197, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_953, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_953, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([1, 2, 12, 256])]; + tensor x_31 = reshape(shape = var_1217, x = xt_3)[name = tensor("x_31")]; + tensor var_1219_perm_0 = const()[name = tensor("op_1219_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1223 = const()[name = tensor("op_1223"), val = tensor([12, 2, 256])]; + tensor var_1219 = transpose(perm = var_1219_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1223, x = var_1219)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1231 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([12, 2, 4, 64])]; + tensor var_1233 = reshape(shape = var_1232, x = var_1231)[name = tensor("op_1233")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1237 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor(0x1p-3)]; + tensor var_1239 = mul(x = var_1237, y = var_1238)[name = tensor("op_1239")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([12, 2, 4, 64])]; + tensor var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor("op_1241")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1245 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1246 = const()[name = tensor("op_1246"), val = tensor([12, 2, 4, 64])]; + tensor var_1247 = reshape(shape = var_1246, x = var_1245)[name = tensor("op_1247")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_958, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1241)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1233)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([2, 1])]; + tensor var_1263 = reshape(shape = var_1262, x = sqrt_s_t)[name = tensor("op_1263")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1263)[name = tensor("M")]; + tensor var_1265 = mul(x = qk, y = M)[name = tensor("op_1265")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1247)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1265, y = v_17)[name = tensor("inner")]; + tensor var_1267_transpose_x_0 = const()[name = tensor("op_1267_transpose_x_0"), val = tensor(false)]; + tensor var_1267_transpose_y_0 = const()[name = tensor("op_1267_transpose_y_0"), val = tensor(false)]; + tensor var_1267 = matmul(transpose_x = var_1267_transpose_x_0, transpose_y = var_1267_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1267")]; + tensor var_1268 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1268")]; + tensor var_1269 = const()[name = tensor("op_1269"), val = tensor([1, 1, 2, 1])]; + tensor var_1270 = reshape(shape = var_1269, x = var_1268)[name = tensor("op_1270")]; + tensor cross = mul(x = var_1267, y = var_1270)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1093)[name = tensor("v_masked")]; + tensor var_1276 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1276")]; + tensor var_1278_transpose_x_1 = const()[name = tensor("op_1278_transpose_x_1"), val = tensor(true)]; + tensor var_1278_transpose_y_1 = const()[name = tensor("op_1278_transpose_y_1"), val = tensor(false)]; + tensor var_1278 = matmul(transpose_x = var_1278_transpose_x_1, transpose_y = var_1278_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1278")]; + tensor new_kv_unnorm = add(x = var_1276, y = var_1278)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1101)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_958, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1287_perm_0 = const()[name = tensor("op_1287_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1287 = transpose(perm = var_1287_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_955, x = var_1287)[name = tensor("out_33")]; + tensor var_1291 = const()[name = tensor("op_1291"), val = tensor([12, 2, 256])]; + tensor out = reshape(shape = var_1291, x = out_33)[name = tensor("out")]; + tensor var_1293 = silu(x = input_187)[name = tensor("op_1293")]; + tensor input_189 = mul(x = var_1293, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_953, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1303 = const()[name = tensor("op_1303"), val = tensor([1, 12, 2, 256])]; + tensor var_1304 = reshape(shape = var_1303, x = xt_5)[name = tensor("op_1304")]; + tensor var_1305_perm_0 = const()[name = tensor("op_1305_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1308 = const()[name = tensor("op_1308"), val = tensor([2, 12, 256])]; + tensor var_1305 = transpose(perm = var_1305_perm_0, x = var_1304)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1308, x = var_1305)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1331 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 2, 3, 256])]; + tensor var_1333 = reshape(shape = concat_2, x = var_1331)[name = tensor("op_1333")]; + tensor var_1334_axes_0 = const()[name = tensor("op_1334_axes_0"), val = tensor([0])]; + tensor var_1334 = expand_dims(axes = var_1334_axes_0, x = var_1333)[name = tensor("op_1334")]; + tensor var_1335_perm_0 = const()[name = tensor("op_1335_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1336_axes_0 = const()[name = tensor("op_1336_axes_0"), val = tensor([-2])]; + tensor var_1335 = transpose(perm = var_1335_perm_0, x = var_1334)[name = tensor("transpose_8")]; + tensor var_1336 = squeeze(axes = var_1336_axes_0, x = var_1335)[name = tensor("op_1336")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 2, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1336)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 2, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1336)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 2, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1336)[name = tensor("v_19")]; + tensor var_1344 = const()[name = tensor("op_1344"), val = tensor([12, 8, 64])]; + tensor var_1345 = reshape(shape = var_1344, x = q_19)[name = tensor("op_1345")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1351 = const()[name = tensor("op_1351"), val = tensor([12, 8, 64])]; + tensor var_1352 = reshape(shape = var_1351, x = k_19)[name = tensor("op_1352")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([12, 8, 64])]; + tensor var_1359 = reshape(shape = var_1358, x = v_19)[name = tensor("op_1359")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1362 = const()[name = tensor("op_1362"), val = tensor([2, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1345)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1362, x = q_21)[name = tensor("q")]; + tensor var_1364 = const()[name = tensor("op_1364"), val = tensor([2, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1352)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1364, x = k_21)[name = tensor("k")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([2, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1359)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1366, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([2, 0, 1, 3])]; + tensor var_1374 = const()[name = tensor("op_1374"), val = tensor([24, 256])]; + tensor var_1370 = transpose(perm = var_1369, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1374, x = var_1370)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1378 = const()[name = tensor("op_1378"), val = tensor([12, 2, 256])]; + tensor attn_output = reshape(shape = var_1378, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_953, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_953, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([1, 2, 12, 256])]; + tensor input = reshape(shape = var_1398, x = xt)[name = tensor("input")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([-1])]; + tensor var_1401 = reduce_l2_norm(axes = var_1400, keep_dims = var_956, x = input)[name = tensor("op_1401")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_948, beta = const_42, x = var_1401)[name = tensor("clip_5")]; + tensor var_1403 = real_div(x = input, y = clip_5)[name = tensor("op_1403")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([2, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([2, 256, 12])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1403)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 2, 12])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 2, 11])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1407")]; + tensor var_1409_axis_0 = const()[name = tensor("op_1409_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1409_axis_0, values = (var_1105, nkv))[name = tensor("op_1409")]; + tensor var_1411_axis_0 = const()[name = tensor("op_1411_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1411_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1411")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/weights/weight.bin b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..f11d3be4c42501ff4528d38b0eb41c7694670dc3 --- /dev/null +++ b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfe0426ee7f10cd239b35836d9c07dbd9406d43ba5cabf2f5584d0a8b7af997f +size 44419968 diff --git a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..c03c36525cece803ff8208a5ea26cac5ac40a745 --- /dev/null +++ b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4835a14c03b80d2ffb221a44b04b7b0b1c4cdb4e6a71682753bcfe838c3773b1 +size 179876 diff --git a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..f11d3be4c42501ff4528d38b0eb41c7694670dc3 --- /dev/null +++ b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfe0426ee7f10cd239b35836d9c07dbd9406d43ba5cabf2f5584d0a8b7af997f +size 44419968 diff --git a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Manifest.json b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..0fd137e70ac537ef68adfd0386c412681e663286 --- /dev/null +++ b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "5BAC2999-7E80-40F0-8220-7BBE6801AB8D": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "ED375499-50AC-4B4A-A792-3AB2C9B91F44": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "ED375499-50AC-4B4A-A792-3AB2C9B91F44" +} diff --git a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/analytics/coremldata.bin b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..65f7eaf6f1cff00f16e9edf9cb4cdfde14a47a56 --- /dev/null +++ b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f1ec2df8f646b9374cef6046fd9256230748f14059ab5f8e8258279b0565bda +size 243 diff --git a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/coremldata.bin b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..10dce50396088c6728d963dc8842a625a7b9facc --- /dev/null +++ b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b37ebc9cf057e7be9a35dd0948213af49004827a34416b9b3634e4609254a179 +size 1310 diff --git a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/metadata.json b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..b80d71250f01b02f0cb7ffcce159276184b82521 --- /dev/null +++ b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND DIHARD III streaming diarizer (pipeline, T=3, max_speakers=10)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 3 × 10)", + "shortDescription" : "", + "shape" : "[1, 3, 10]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 68, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 56, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 18, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 47 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 3 × 345)", + "shortDescription" : "", + "shape" : "[1, 3, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 3)", + "shortDescription" : "", + "shape" : "[3]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"dih3\", \"model_label\": \"DIHARD III\", \"variant\": \"pipeline\", \"chunk_size\": 3, \"step_duration_ms\": 300, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_dih3_300ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/model.mil b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..48f1c415ceeabfb14a7edfb748becb4366d6de9f --- /dev/null +++ b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/model.mil @@ -0,0 +1,1286 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 3, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 3, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 3, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([3, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 3, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1.8p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 3, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; + tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; + tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; + tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; + tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; + tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; + tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; + tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; + tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; + tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; + tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; + tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; + tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; + tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 1, 256])]; + tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, true, true])]; + tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; + tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; + tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; + tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; + tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7))[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_265_split_sizes_0 = const()[name = tensor("op_265_split_sizes_0"), val = tensor([256, 256])]; + tensor var_265_axis_0 = const()[name = tensor("op_265_axis_0"), val = tensor(1)]; + tensor var_265_0, tensor var_265_1 = split(axis = var_265_axis_0, split_sizes = var_265_split_sizes_0, x = inputs_3)[name = tensor("op_265")]; + tensor var_267 = sigmoid(x = var_265_1)[name = tensor("op_267")]; + tensor inputs_5 = mul(x = var_265_0, y = var_267)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([3, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_298_begin_0 = const()[name = tensor("op_298_begin_0"), val = tensor([0, -1, 0])]; + tensor var_298_end_0 = const()[name = tensor("op_298_end_0"), val = tensor([3, 16, 256])]; + tensor var_298_end_mask_0 = const()[name = tensor("op_298_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_298 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, x = conv_out_1)[name = tensor("op_298")]; + tensor var_300_perm_0 = const()[name = tensor("op_300_perm_0"), val = tensor([1, 0, 2])]; + tensor var_300 = transpose(perm = var_300_perm_0, x = var_298)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_300)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_323 = const()[name = tensor("op_323"), val = tensor(0x1p-1)]; + tensor var_324 = mul(x = input_39, y = var_323)[name = tensor("op_324")]; + tensor input_41 = add(x = var_324, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_353 = const()[name = tensor("op_353"), val = tensor(0x1p-1)]; + tensor var_354 = mul(x = input_51, y = var_353)[name = tensor("op_354")]; + tensor input_53 = add(x = var_354, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_368 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 3, 4, 64])]; + tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_374 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_375 = const()[name = tensor("op_375"), val = tensor(0x1p-3)]; + tensor var_376 = mul(x = var_374, y = var_375)[name = tensor("op_376")]; + tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 3, 4, 64])]; + tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_382 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_383 = const()[name = tensor("op_383"), val = tensor([1, 3, 4, 64])]; + tensor var_384 = reshape(shape = var_383, x = var_382)[name = tensor("op_384")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_378)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_370)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_394 = const()[name = tensor("op_394"), val = tensor([3, 1])]; + tensor var_395 = reshape(shape = var_394, x = sqrt_s_t_3)[name = tensor("op_395")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_395)[name = tensor("M_3")]; + tensor var_397 = mul(x = qk_3, y = M_3)[name = tensor("op_397")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_384)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_397, y = v_3)[name = tensor("inner_3")]; + tensor var_399_transpose_x_0 = const()[name = tensor("op_399_transpose_x_0"), val = tensor(false)]; + tensor var_399_transpose_y_0 = const()[name = tensor("op_399_transpose_y_0"), val = tensor(false)]; + tensor var_399 = matmul(transpose_x = var_399_transpose_x_0, transpose_y = var_399_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_399")]; + tensor var_400 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_400")]; + tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 3, 1])]; + tensor var_402 = reshape(shape = var_401, x = var_400)[name = tensor("op_402")]; + tensor cross_3 = mul(x = var_399, y = var_402)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_405 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_405")]; + tensor var_407_transpose_x_1 = const()[name = tensor("op_407_transpose_x_1"), val = tensor(true)]; + tensor var_407_transpose_y_1 = const()[name = tensor("op_407_transpose_y_1"), val = tensor(false)]; + tensor var_407 = matmul(transpose_x = var_407_transpose_x_1, transpose_y = var_407_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_407")]; + tensor new_kv_unnorm_3 = add(x = var_405, y = var_407)[name = tensor("new_kv_unnorm_3")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor(0x1.8p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_409)[name = tensor("new_scale_3")]; + tensor var_411 = sqrt(x = new_scale_3)[name = tensor("op_411")]; + tensor var_412 = real_div(x = new_kv_unnorm_3, y = var_411)[name = tensor("op_412")]; + tensor var_413_perm_0 = const()[name = tensor("op_413_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_413 = transpose(perm = var_413_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_413)[name = tensor("out_9")]; + tensor var_417 = const()[name = tensor("op_417"), val = tensor([1, 3, 256])]; + tensor out_11 = reshape(shape = var_417, x = out_9)[name = tensor("out_11")]; + tensor var_419 = silu(x = input_57)[name = tensor("op_419")]; + tensor input_59 = mul(x = var_419, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; + tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 0, 0])]; + tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; + tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, false, true])]; + tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; + tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; + tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; + tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; + tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; + tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 1, 0])]; + tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 2, 256])]; + tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; + tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; + tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; + tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; + tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; + tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; + tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; + tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 2, 0])]; + tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 1, 256])]; + tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, true, true])]; + tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; + tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; + tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; + tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; + tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_11, window_13, window_15))[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_471_split_sizes_0 = const()[name = tensor("op_471_split_sizes_0"), val = tensor([256, 256])]; + tensor var_471_axis_0 = const()[name = tensor("op_471_axis_0"), val = tensor(1)]; + tensor var_471_0, tensor var_471_1 = split(axis = var_471_axis_0, split_sizes = var_471_split_sizes_0, x = inputs_13)[name = tensor("op_471")]; + tensor var_473 = sigmoid(x = var_471_1)[name = tensor("op_473")]; + tensor inputs_15 = mul(x = var_471_0, y = var_473)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([3, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_504_begin_0 = const()[name = tensor("op_504_begin_0"), val = tensor([0, -1, 0])]; + tensor var_504_end_0 = const()[name = tensor("op_504_end_0"), val = tensor([3, 16, 256])]; + tensor var_504_end_mask_0 = const()[name = tensor("op_504_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_504 = slice_by_index(begin = var_504_begin_0, end = var_504_end_0, end_mask = var_504_end_mask_0, x = conv_out_3)[name = tensor("op_504")]; + tensor var_506_perm_0 = const()[name = tensor("op_506_perm_0"), val = tensor([1, 0, 2])]; + tensor var_506 = transpose(perm = var_506_perm_0, x = var_504)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_506)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_529 = const()[name = tensor("op_529"), val = tensor(0x1p-1)]; + tensor var_530 = mul(x = input_79, y = var_529)[name = tensor("op_530")]; + tensor input_81 = add(x = var_530, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_559 = const()[name = tensor("op_559"), val = tensor(0x1p-1)]; + tensor var_560 = mul(x = input_91, y = var_559)[name = tensor("op_560")]; + tensor input_93 = add(x = var_560, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_574 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 3, 4, 64])]; + tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_580 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_581 = const()[name = tensor("op_581"), val = tensor(0x1p-3)]; + tensor var_582 = mul(x = var_580, y = var_581)[name = tensor("op_582")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 3, 4, 64])]; + tensor var_584 = reshape(shape = var_583, x = var_582)[name = tensor("op_584")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_588 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_589 = const()[name = tensor("op_589"), val = tensor([1, 3, 4, 64])]; + tensor var_590 = reshape(shape = var_589, x = var_588)[name = tensor("op_590")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_584)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_576)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_600 = const()[name = tensor("op_600"), val = tensor([3, 1])]; + tensor var_601 = reshape(shape = var_600, x = sqrt_s_t_5)[name = tensor("op_601")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_601)[name = tensor("M_5")]; + tensor var_603 = mul(x = qk_5, y = M_5)[name = tensor("op_603")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_590)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_603, y = v_5)[name = tensor("inner_5")]; + tensor var_605_transpose_x_0 = const()[name = tensor("op_605_transpose_x_0"), val = tensor(false)]; + tensor var_605_transpose_y_0 = const()[name = tensor("op_605_transpose_y_0"), val = tensor(false)]; + tensor var_605 = matmul(transpose_x = var_605_transpose_x_0, transpose_y = var_605_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_605")]; + tensor var_606 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_606")]; + tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 1, 3, 1])]; + tensor var_608 = reshape(shape = var_607, x = var_606)[name = tensor("op_608")]; + tensor cross_5 = mul(x = var_605, y = var_608)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_611 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_611")]; + tensor var_613_transpose_x_1 = const()[name = tensor("op_613_transpose_x_1"), val = tensor(true)]; + tensor var_613_transpose_y_1 = const()[name = tensor("op_613_transpose_y_1"), val = tensor(false)]; + tensor var_613 = matmul(transpose_x = var_613_transpose_x_1, transpose_y = var_613_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_613")]; + tensor new_kv_unnorm_5 = add(x = var_611, y = var_613)[name = tensor("new_kv_unnorm_5")]; + tensor var_615 = const()[name = tensor("op_615"), val = tensor(0x1.8p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_615)[name = tensor("new_scale_5")]; + tensor var_617 = sqrt(x = new_scale_5)[name = tensor("op_617")]; + tensor var_618 = real_div(x = new_kv_unnorm_5, y = var_617)[name = tensor("op_618")]; + tensor var_619_perm_0 = const()[name = tensor("op_619_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_619 = transpose(perm = var_619_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_619)[name = tensor("out_15")]; + tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 3, 256])]; + tensor out_17 = reshape(shape = var_623, x = out_15)[name = tensor("out_17")]; + tensor var_625 = silu(x = input_97)[name = tensor("op_625")]; + tensor input_99 = mul(x = var_625, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_17_begin_0 = const()[name = tensor("window_17_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_17_end_0 = const()[name = tensor("window_17_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_17_end_mask_0 = const()[name = tensor("window_17_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_17_squeeze_mask_0 = const()[name = tensor("window_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_17 = slice_by_index(begin = window_17_begin_0, end = window_17_end_0, end_mask = window_17_end_mask_0, squeeze_mask = window_17_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_17")]; + tensor var_633_begin_0 = const()[name = tensor("op_633_begin_0"), val = tensor([0, 0, 0])]; + tensor var_633_end_0 = const()[name = tensor("op_633_end_0"), val = tensor([1, 1, 256])]; + tensor var_633_end_mask_0 = const()[name = tensor("op_633_end_mask_0"), val = tensor([true, false, true])]; + tensor var_633 = slice_by_index(begin = var_633_begin_0, end = var_633_end_0, end_mask = var_633_end_mask_0, x = x_15)[name = tensor("op_633")]; + tensor var_636_begin_0 = const()[name = tensor("op_636_begin_0"), val = tensor([0, 1, 0])]; + tensor var_636_end_0 = const()[name = tensor("op_636_end_0"), val = tensor([1, 16, 256])]; + tensor var_636_end_mask_0 = const()[name = tensor("op_636_end_mask_0"), val = tensor([true, true, true])]; + tensor var_636 = slice_by_index(begin = var_636_begin_0, end = var_636_end_0, end_mask = var_636_end_mask_0, x = window_17)[name = tensor("op_636")]; + tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; + tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_636, var_633))[name = tensor("window_19")]; + tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 1, 0])]; + tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 2, 256])]; + tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, false, true])]; + tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = x_15)[name = tensor("op_641")]; + tensor var_644_begin_0 = const()[name = tensor("op_644_begin_0"), val = tensor([0, 1, 0])]; + tensor var_644_end_0 = const()[name = tensor("op_644_end_0"), val = tensor([1, 16, 256])]; + tensor var_644_end_mask_0 = const()[name = tensor("op_644_end_mask_0"), val = tensor([true, true, true])]; + tensor var_644 = slice_by_index(begin = var_644_begin_0, end = var_644_end_0, end_mask = var_644_end_mask_0, x = window_19)[name = tensor("op_644")]; + tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; + tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_644, var_641))[name = tensor("window_21")]; + tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 2, 0])]; + tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; + tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, true, true])]; + tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; + tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; + tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; + tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; + tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; + tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_19, window_21, window_23))[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_677_split_sizes_0 = const()[name = tensor("op_677_split_sizes_0"), val = tensor([256, 256])]; + tensor var_677_axis_0 = const()[name = tensor("op_677_axis_0"), val = tensor(1)]; + tensor var_677_0, tensor var_677_1 = split(axis = var_677_axis_0, split_sizes = var_677_split_sizes_0, x = inputs_23)[name = tensor("op_677")]; + tensor var_679 = sigmoid(x = var_677_1)[name = tensor("op_679")]; + tensor inputs_25 = mul(x = var_677_0, y = var_679)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([3, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_710_begin_0 = const()[name = tensor("op_710_begin_0"), val = tensor([0, -1, 0])]; + tensor var_710_end_0 = const()[name = tensor("op_710_end_0"), val = tensor([3, 16, 256])]; + tensor var_710_end_mask_0 = const()[name = tensor("op_710_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_710 = slice_by_index(begin = var_710_begin_0, end = var_710_end_0, end_mask = var_710_end_mask_0, x = conv_out_5)[name = tensor("op_710")]; + tensor var_712_perm_0 = const()[name = tensor("op_712_perm_0"), val = tensor([1, 0, 2])]; + tensor var_712 = transpose(perm = var_712_perm_0, x = var_710)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_712)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_735 = const()[name = tensor("op_735"), val = tensor(0x1p-1)]; + tensor var_736 = mul(x = input_119, y = var_735)[name = tensor("op_736")]; + tensor input_121 = add(x = var_736, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor(0x1p-1)]; + tensor var_766 = mul(x = input_131, y = var_765)[name = tensor("op_766")]; + tensor input_133 = add(x = var_766, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_780 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 3, 4, 64])]; + tensor var_782 = reshape(shape = var_781, x = var_780)[name = tensor("op_782")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_786 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_787 = const()[name = tensor("op_787"), val = tensor(0x1p-3)]; + tensor var_788 = mul(x = var_786, y = var_787)[name = tensor("op_788")]; + tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 3, 4, 64])]; + tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_794 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_795 = const()[name = tensor("op_795"), val = tensor([1, 3, 4, 64])]; + tensor var_796 = reshape(shape = var_795, x = var_794)[name = tensor("op_796")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_790)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_782)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_806 = const()[name = tensor("op_806"), val = tensor([3, 1])]; + tensor var_807 = reshape(shape = var_806, x = sqrt_s_t_7)[name = tensor("op_807")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_807)[name = tensor("M_7")]; + tensor var_809 = mul(x = qk_7, y = M_7)[name = tensor("op_809")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_796)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_809, y = v_7)[name = tensor("inner_7")]; + tensor var_811_transpose_x_0 = const()[name = tensor("op_811_transpose_x_0"), val = tensor(false)]; + tensor var_811_transpose_y_0 = const()[name = tensor("op_811_transpose_y_0"), val = tensor(false)]; + tensor var_811 = matmul(transpose_x = var_811_transpose_x_0, transpose_y = var_811_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_811")]; + tensor var_812 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_812")]; + tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 1, 3, 1])]; + tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; + tensor cross_7 = mul(x = var_811, y = var_814)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_817 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_817")]; + tensor var_819_transpose_x_1 = const()[name = tensor("op_819_transpose_x_1"), val = tensor(true)]; + tensor var_819_transpose_y_1 = const()[name = tensor("op_819_transpose_y_1"), val = tensor(false)]; + tensor var_819 = matmul(transpose_x = var_819_transpose_x_1, transpose_y = var_819_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_819")]; + tensor new_kv_unnorm_7 = add(x = var_817, y = var_819)[name = tensor("new_kv_unnorm_7")]; + tensor var_821 = const()[name = tensor("op_821"), val = tensor(0x1.8p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_821)[name = tensor("new_scale_7")]; + tensor var_823 = sqrt(x = new_scale_7)[name = tensor("op_823")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_823)[name = tensor("nkv_1")]; + tensor var_825_perm_0 = const()[name = tensor("op_825_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_825 = transpose(perm = var_825_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_825)[name = tensor("out_21")]; + tensor var_829 = const()[name = tensor("op_829"), val = tensor([1, 3, 256])]; + tensor out_23 = reshape(shape = var_829, x = out_21)[name = tensor("out_23")]; + tensor var_831 = silu(x = input_137)[name = tensor("op_831")]; + tensor input_139 = mul(x = var_831, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; + tensor var_839_begin_0 = const()[name = tensor("op_839_begin_0"), val = tensor([0, 0, 0])]; + tensor var_839_end_0 = const()[name = tensor("op_839_end_0"), val = tensor([1, 1, 256])]; + tensor var_839_end_mask_0 = const()[name = tensor("op_839_end_mask_0"), val = tensor([true, false, true])]; + tensor var_839 = slice_by_index(begin = var_839_begin_0, end = var_839_end_0, end_mask = var_839_end_mask_0, x = x_21)[name = tensor("op_839")]; + tensor var_842_begin_0 = const()[name = tensor("op_842_begin_0"), val = tensor([0, 1, 0])]; + tensor var_842_end_0 = const()[name = tensor("op_842_end_0"), val = tensor([1, 16, 256])]; + tensor var_842_end_mask_0 = const()[name = tensor("op_842_end_mask_0"), val = tensor([true, true, true])]; + tensor var_842 = slice_by_index(begin = var_842_begin_0, end = var_842_end_0, end_mask = var_842_end_mask_0, x = window_25)[name = tensor("op_842")]; + tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; + tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_842, var_839))[name = tensor("window_27")]; + tensor var_847_begin_0 = const()[name = tensor("op_847_begin_0"), val = tensor([0, 1, 0])]; + tensor var_847_end_0 = const()[name = tensor("op_847_end_0"), val = tensor([1, 2, 256])]; + tensor var_847_end_mask_0 = const()[name = tensor("op_847_end_mask_0"), val = tensor([true, false, true])]; + tensor var_847 = slice_by_index(begin = var_847_begin_0, end = var_847_end_0, end_mask = var_847_end_mask_0, x = x_21)[name = tensor("op_847")]; + tensor var_850_begin_0 = const()[name = tensor("op_850_begin_0"), val = tensor([0, 1, 0])]; + tensor var_850_end_0 = const()[name = tensor("op_850_end_0"), val = tensor([1, 16, 256])]; + tensor var_850_end_mask_0 = const()[name = tensor("op_850_end_mask_0"), val = tensor([true, true, true])]; + tensor var_850 = slice_by_index(begin = var_850_begin_0, end = var_850_end_0, end_mask = var_850_end_mask_0, x = window_27)[name = tensor("op_850")]; + tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; + tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_850, var_847))[name = tensor("window_29")]; + tensor var_855_begin_0 = const()[name = tensor("op_855_begin_0"), val = tensor([0, 2, 0])]; + tensor var_855_end_0 = const()[name = tensor("op_855_end_0"), val = tensor([1, 1, 256])]; + tensor var_855_end_mask_0 = const()[name = tensor("op_855_end_mask_0"), val = tensor([true, true, true])]; + tensor var_855 = slice_by_index(begin = var_855_begin_0, end = var_855_end_0, end_mask = var_855_end_mask_0, x = x_21)[name = tensor("op_855")]; + tensor var_858_begin_0 = const()[name = tensor("op_858_begin_0"), val = tensor([0, 1, 0])]; + tensor var_858_end_0 = const()[name = tensor("op_858_end_0"), val = tensor([1, 16, 256])]; + tensor var_858_end_mask_0 = const()[name = tensor("op_858_end_mask_0"), val = tensor([true, true, true])]; + tensor var_858 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = window_29)[name = tensor("op_858")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_858, var_855))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_27, window_29, window))[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_883_split_sizes_0 = const()[name = tensor("op_883_split_sizes_0"), val = tensor([256, 256])]; + tensor var_883_axis_0 = const()[name = tensor("op_883_axis_0"), val = tensor(1)]; + tensor var_883_0, tensor var_883_1 = split(axis = var_883_axis_0, split_sizes = var_883_split_sizes_0, x = inputs_33)[name = tensor("op_883")]; + tensor var_885 = sigmoid(x = var_883_1)[name = tensor("op_885")]; + tensor inputs_35 = mul(x = var_883_0, y = var_885)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([3, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_916_begin_0 = const()[name = tensor("op_916_begin_0"), val = tensor([0, -1, 0])]; + tensor var_916_end_0 = const()[name = tensor("op_916_end_0"), val = tensor([3, 16, 256])]; + tensor var_916_end_mask_0 = const()[name = tensor("op_916_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_916 = slice_by_index(begin = var_916_begin_0, end = var_916_end_0, end_mask = var_916_end_mask_0, x = conv_out_7)[name = tensor("op_916")]; + tensor var_918_perm_0 = const()[name = tensor("op_918_perm_0"), val = tensor([1, 0, 2])]; + tensor var_918 = transpose(perm = var_918_perm_0, x = var_916)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_918)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_941 = const()[name = tensor("op_941"), val = tensor(0x1p-1)]; + tensor var_942 = mul(x = input_159, y = var_941)[name = tensor("op_942")]; + tensor input_161 = add(x = var_942, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_960_begin_0 = const()[name = tensor("op_960_begin_0"), val = tensor([0, 0, 3])]; + tensor var_960_end_0 = const()[name = tensor("op_960_end_0"), val = tensor([1, 256, 21])]; + tensor var_960_end_mask_0 = const()[name = tensor("op_960_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_960_begin_0, end = var_960_end_0, end_mask = var_960_end_mask_0, x = cat)[name = tensor("op_960")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_963 = reduce_l2_norm(axes = var_962, keep_dims = var_29, x = input_163)[name = tensor("op_963")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_963)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_967_axis_0, values = (var_206, var_412, var_618, nkv_1))[name = tensor("op_967")]; + tensor var_969_axis_0 = const()[name = tensor("op_969_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_969_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_969")]; + tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_971_axis_0, values = (window_7, window_15, window_23, window))[name = tensor("op_971")]; + tensor var_980 = const()[name = tensor("op_980"), val = tensor(0x1.5798eep-27)]; + tensor var_985 = const()[name = tensor("op_985"), val = tensor(0x1.4f8b58p-17)]; + tensor var_987 = const()[name = tensor("op_987"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_988 = const()[name = tensor("op_988"), val = tensor(true)]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor(0x1p+0)]; + tensor var_994 = const()[name = tensor("op_994"), val = tensor(-1)]; + tensor var_1000 = const()[name = tensor("op_1000"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_1062_axes_0 = const()[name = tensor("op_1062_axes_0"), val = tensor([2])]; + tensor var_1062 = expand_dims(axes = var_1062_axes_0, x = emb)[name = tensor("op_1062")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1062)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_994, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1070_perm_0 = const()[name = tensor("op_1070_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1074 = const()[name = tensor("op_1074"), val = tensor([12, 3, 256])]; + tensor var_1070 = transpose(perm = var_1070_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1074, x = var_1070)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1082 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1083 = const()[name = tensor("op_1083"), val = tensor([12, 3, 4, 64])]; + tensor var_1084 = reshape(shape = var_1083, x = var_1082)[name = tensor("op_1084")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1088 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1089 = const()[name = tensor("op_1089"), val = tensor(0x1p-3)]; + tensor var_1090 = mul(x = var_1088, y = var_1089)[name = tensor("op_1090")]; + tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([12, 3, 4, 64])]; + tensor var_1092 = reshape(shape = var_1091, x = var_1090)[name = tensor("op_1092")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1096 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([12, 3, 4, 64])]; + tensor var_1098 = reshape(shape = var_1097, x = var_1096)[name = tensor("op_1098")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_1000, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_990, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1092)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1084)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([1, 3])]; + tensor var_1111 = reshape(shape = var_1110, x = valid_mask)[name = tensor("op_1111")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1111)[name = tensor("causal_with_valid_1")]; + tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([3, 1])]; + tensor var_1114 = reshape(shape = var_1113, x = sqrt_s_t_9)[name = tensor("op_1114")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1114)[name = tensor("M_9")]; + tensor var_1116 = mul(x = qk_9, y = M_9)[name = tensor("op_1116")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1098)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1116, y = v_9)[name = tensor("inner_9")]; + tensor var_1118_transpose_x_0 = const()[name = tensor("op_1118_transpose_x_0"), val = tensor(false)]; + tensor var_1118_transpose_y_0 = const()[name = tensor("op_1118_transpose_y_0"), val = tensor(false)]; + tensor var_1118 = matmul(transpose_x = var_1118_transpose_x_0, transpose_y = var_1118_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1118")]; + tensor var_1119 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1119")]; + tensor var_1120 = const()[name = tensor("op_1120"), val = tensor([1, 1, 3, 1])]; + tensor var_1121 = reshape(shape = var_1120, x = var_1119)[name = tensor("op_1121")]; + tensor cross_9 = mul(x = var_1118, y = var_1121)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1124 = const()[name = tensor("op_1124"), val = tensor([1, 1, 3, 1])]; + tensor var_1125 = reshape(shape = var_1124, x = valid_mask)[name = tensor("op_1125")]; + tensor v_masked_1 = mul(x = v_9, y = var_1125)[name = tensor("v_masked_1")]; + tensor var_1127 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1127")]; + tensor var_1129_transpose_x_1 = const()[name = tensor("op_1129_transpose_x_1"), val = tensor(true)]; + tensor var_1129_transpose_y_1 = const()[name = tensor("op_1129_transpose_y_1"), val = tensor(false)]; + tensor var_1129 = matmul(transpose_x = var_1129_transpose_x_1, transpose_y = var_1129_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1129")]; + tensor new_kv_unnorm_9 = add(x = var_1127, y = var_1129)[name = tensor("new_kv_unnorm_9")]; + tensor var_1131_keep_dims_0 = const()[name = tensor("op_1131_keep_dims_0"), val = tensor(false)]; + tensor var_1131 = reduce_sum(keep_dims = var_1131_keep_dims_0, x = valid_mask)[name = tensor("op_1131")]; + tensor var_1132 = const()[name = tensor("op_1132"), val = tensor([1])]; + tensor var_1133 = reshape(shape = var_1132, x = var_1131)[name = tensor("op_1133")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1133)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_990, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1137 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1137")]; + tensor var_1138_perm_0 = const()[name = tensor("op_1138_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1138 = transpose(perm = var_1138_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_987, x = var_1138)[name = tensor("out_27")]; + tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([12, 3, 256])]; + tensor out_29 = reshape(shape = var_1142, x = out_27)[name = tensor("out_29")]; + tensor var_1144 = silu(x = input_169)[name = tensor("op_1144")]; + tensor input_171 = mul(x = var_1144, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_985, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 12, 3, 256])]; + tensor var_1155 = reshape(shape = var_1154, x = xt_1)[name = tensor("op_1155")]; + tensor var_1156_perm_0 = const()[name = tensor("op_1156_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1159 = const()[name = tensor("op_1159"), val = tensor([3, 12, 256])]; + tensor var_1156 = transpose(perm = var_1156_perm_0, x = var_1155)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1159, x = var_1156)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1182 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 3, 3, 256])]; + tensor var_1184 = reshape(shape = concat_1, x = var_1182)[name = tensor("op_1184")]; + tensor var_1185_axes_0 = const()[name = tensor("op_1185_axes_0"), val = tensor([0])]; + tensor var_1185 = expand_dims(axes = var_1185_axes_0, x = var_1184)[name = tensor("op_1185")]; + tensor var_1186_perm_0 = const()[name = tensor("op_1186_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1187_axes_0 = const()[name = tensor("op_1187_axes_0"), val = tensor([-2])]; + tensor var_1186 = transpose(perm = var_1186_perm_0, x = var_1185)[name = tensor("transpose_21")]; + tensor var_1187 = squeeze(axes = var_1187_axes_0, x = var_1186)[name = tensor("op_1187")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 3, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1187)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 3, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1187)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 3, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1187)[name = tensor("v_11")]; + tensor var_1195 = const()[name = tensor("op_1195"), val = tensor([12, 12, 64])]; + tensor var_1196 = reshape(shape = var_1195, x = q_11)[name = tensor("op_1196")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1202 = const()[name = tensor("op_1202"), val = tensor([12, 12, 64])]; + tensor var_1203 = reshape(shape = var_1202, x = k_11)[name = tensor("op_1203")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1209 = const()[name = tensor("op_1209"), val = tensor([12, 12, 64])]; + tensor var_1210 = reshape(shape = var_1209, x = v_11)[name = tensor("op_1210")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1213 = const()[name = tensor("op_1213"), val = tensor([3, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1196)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1213, x = q_13)[name = tensor("q_15")]; + tensor var_1215 = const()[name = tensor("op_1215"), val = tensor([3, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1203)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1215, x = k_13)[name = tensor("k_15")]; + tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([3, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1210)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1217, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1220 = const()[name = tensor("op_1220"), val = tensor([2, 0, 1, 3])]; + tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([36, 256])]; + tensor var_1221 = transpose(perm = var_1220, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1225, x = var_1221)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1229 = const()[name = tensor("op_1229"), val = tensor([12, 3, 256])]; + tensor attn_output_7 = reshape(shape = var_1229, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_985, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_985, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([1, 3, 12, 256])]; + tensor x_31 = reshape(shape = var_1249, x = xt_3)[name = tensor("x_31")]; + tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1255 = const()[name = tensor("op_1255"), val = tensor([12, 3, 256])]; + tensor var_1251 = transpose(perm = var_1251_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1255, x = var_1251)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1263 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1264 = const()[name = tensor("op_1264"), val = tensor([12, 3, 4, 64])]; + tensor var_1265 = reshape(shape = var_1264, x = var_1263)[name = tensor("op_1265")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1269 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1270 = const()[name = tensor("op_1270"), val = tensor(0x1p-3)]; + tensor var_1271 = mul(x = var_1269, y = var_1270)[name = tensor("op_1271")]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([12, 3, 4, 64])]; + tensor var_1273 = reshape(shape = var_1272, x = var_1271)[name = tensor("op_1273")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1277 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([12, 3, 4, 64])]; + tensor var_1279 = reshape(shape = var_1278, x = var_1277)[name = tensor("op_1279")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_990, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1273)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1265)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([3, 1])]; + tensor var_1295 = reshape(shape = var_1294, x = sqrt_s_t)[name = tensor("op_1295")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1295)[name = tensor("M")]; + tensor var_1297 = mul(x = qk, y = M)[name = tensor("op_1297")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1279)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1297, y = v_17)[name = tensor("inner")]; + tensor var_1299_transpose_x_0 = const()[name = tensor("op_1299_transpose_x_0"), val = tensor(false)]; + tensor var_1299_transpose_y_0 = const()[name = tensor("op_1299_transpose_y_0"), val = tensor(false)]; + tensor var_1299 = matmul(transpose_x = var_1299_transpose_x_0, transpose_y = var_1299_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1299")]; + tensor var_1300 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1300")]; + tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([1, 1, 3, 1])]; + tensor var_1302 = reshape(shape = var_1301, x = var_1300)[name = tensor("op_1302")]; + tensor cross = mul(x = var_1299, y = var_1302)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1125)[name = tensor("v_masked")]; + tensor var_1308 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1308")]; + tensor var_1310_transpose_x_1 = const()[name = tensor("op_1310_transpose_x_1"), val = tensor(true)]; + tensor var_1310_transpose_y_1 = const()[name = tensor("op_1310_transpose_y_1"), val = tensor(false)]; + tensor var_1310 = matmul(transpose_x = var_1310_transpose_x_1, transpose_y = var_1310_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1310")]; + tensor new_kv_unnorm = add(x = var_1308, y = var_1310)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1133)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_990, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1319_perm_0 = const()[name = tensor("op_1319_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1319 = transpose(perm = var_1319_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_987, x = var_1319)[name = tensor("out_33")]; + tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([12, 3, 256])]; + tensor out = reshape(shape = var_1323, x = out_33)[name = tensor("out")]; + tensor var_1325 = silu(x = input_187)[name = tensor("op_1325")]; + tensor input_189 = mul(x = var_1325, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_985, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor([1, 12, 3, 256])]; + tensor var_1336 = reshape(shape = var_1335, x = xt_5)[name = tensor("op_1336")]; + tensor var_1337_perm_0 = const()[name = tensor("op_1337_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1340 = const()[name = tensor("op_1340"), val = tensor([3, 12, 256])]; + tensor var_1337 = transpose(perm = var_1337_perm_0, x = var_1336)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1340, x = var_1337)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1363 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 3, 3, 256])]; + tensor var_1365 = reshape(shape = concat_2, x = var_1363)[name = tensor("op_1365")]; + tensor var_1366_axes_0 = const()[name = tensor("op_1366_axes_0"), val = tensor([0])]; + tensor var_1366 = expand_dims(axes = var_1366_axes_0, x = var_1365)[name = tensor("op_1366")]; + tensor var_1367_perm_0 = const()[name = tensor("op_1367_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1368_axes_0 = const()[name = tensor("op_1368_axes_0"), val = tensor([-2])]; + tensor var_1367 = transpose(perm = var_1367_perm_0, x = var_1366)[name = tensor("transpose_8")]; + tensor var_1368 = squeeze(axes = var_1368_axes_0, x = var_1367)[name = tensor("op_1368")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 3, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1368)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 3, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1368)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 3, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1368)[name = tensor("v_19")]; + tensor var_1376 = const()[name = tensor("op_1376"), val = tensor([12, 12, 64])]; + tensor var_1377 = reshape(shape = var_1376, x = q_19)[name = tensor("op_1377")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([12, 12, 64])]; + tensor var_1384 = reshape(shape = var_1383, x = k_19)[name = tensor("op_1384")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([12, 12, 64])]; + tensor var_1391 = reshape(shape = var_1390, x = v_19)[name = tensor("op_1391")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1394 = const()[name = tensor("op_1394"), val = tensor([3, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1377)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1394, x = q_21)[name = tensor("q")]; + tensor var_1396 = const()[name = tensor("op_1396"), val = tensor([3, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1384)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1396, x = k_21)[name = tensor("k")]; + tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([3, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1391)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1398, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1401 = const()[name = tensor("op_1401"), val = tensor([2, 0, 1, 3])]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([36, 256])]; + tensor var_1402 = transpose(perm = var_1401, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1406, x = var_1402)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1410 = const()[name = tensor("op_1410"), val = tensor([12, 3, 256])]; + tensor attn_output = reshape(shape = var_1410, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_985, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_985, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([1, 3, 12, 256])]; + tensor input = reshape(shape = var_1430, x = xt)[name = tensor("input")]; + tensor var_1432 = const()[name = tensor("op_1432"), val = tensor([-1])]; + tensor var_1433 = reduce_l2_norm(axes = var_1432, keep_dims = var_988, x = input)[name = tensor("op_1433")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_980, beta = const_42, x = var_1433)[name = tensor("clip_5")]; + tensor var_1435 = real_div(x = input, y = clip_5)[name = tensor("op_1435")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([3, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([3, 256, 12])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1435)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 3, 12])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 3, 11])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1439")]; + tensor var_1441_axis_0 = const()[name = tensor("op_1441_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1441_axis_0, values = (var_1137, nkv))[name = tensor("op_1441")]; + tensor var_1443_axis_0 = const()[name = tensor("op_1443_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1443_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1443")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/weights/weight.bin b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..8099728777142c50cf4033f9aca1e7c76b338438 --- /dev/null +++ b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:460cb40cf6b38e8edb340fbc53c8182edb442e992f4a1cb2fb530121f4b9a889 +size 44432256 diff --git a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..d17264963ac53e18b1094bfefb69b97ef1395517 --- /dev/null +++ b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9986be1fe8f5022bafee300a2d8b49effe1540945ddb6ae9eaab94821982c313 +size 185469 diff --git a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..8099728777142c50cf4033f9aca1e7c76b338438 --- /dev/null +++ b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:460cb40cf6b38e8edb340fbc53c8182edb442e992f4a1cb2fb530121f4b9a889 +size 44432256 diff --git a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Manifest.json b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d174cfd424c10645eda95ea853dbe518047303b0 --- /dev/null +++ b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "5F03C3B0-3340-4DD2-B75B-F78D5596775F": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "F681008A-C1D4-4681-92CB-73FB966EF8F3": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "F681008A-C1D4-4681-92CB-73FB966EF8F3" +} diff --git a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/analytics/coremldata.bin b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..71fa07880e9abdfe454605d65a2e99de84ec0b8c --- /dev/null +++ b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88851cb773f9d1d6b36189e9c262dcc461c0d873a69a1f1dcf4087f48c4dc60b +size 243 diff --git a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/coremldata.bin b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..18f01273657c24b054f38d6280b09122f7253d2a --- /dev/null +++ b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:606fe839f02edeb537028ed20f1abdc784cf5a18cef58ae9ff1a2783540da01b +size 1310 diff --git a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/metadata.json b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..cba5d30a2ad995503a2fcbad94a71f5afd760118 --- /dev/null +++ b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND DIHARD III streaming diarizer (pipeline, T=4, max_speakers=10)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 4 × 10)", + "shortDescription" : "", + "shape" : "[1, 4, 10]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 68, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 64, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 22, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 47 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 4 × 345)", + "shortDescription" : "", + "shape" : "[1, 4, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4)", + "shortDescription" : "", + "shape" : "[4]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"dih3\", \"model_label\": \"DIHARD III\", \"variant\": \"pipeline\", \"chunk_size\": 4, \"step_duration_ms\": 400, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_dih3_400ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/model.mil b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..7135f6328b831d1038e863b851b7563d75bba5c3 --- /dev/null +++ b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/model.mil @@ -0,0 +1,1326 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982080)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983168)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336512)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337600)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338688)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339776)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340864)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345024)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393664)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394752)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443392)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444480)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445568)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446656)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708864)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709952)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972160)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973248)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235456)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236544)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498752)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499840)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762048)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763136)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764224)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766336)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290688)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307136)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308224)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309312)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310400)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311488)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312576)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574784)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575872)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576960)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581120)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629760)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630848)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679488)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680576)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681664)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682752)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683840)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688000)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736640)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737728)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786368)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787456)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788544)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789632)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051840)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052928)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315136)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316224)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578432)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579520)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841728)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842816)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105024)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106112)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107200)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109312)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633664)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650112)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651200)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652288)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653376)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654464)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655552)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917760)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918848)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919936)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924096)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972736)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973824)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022464)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023552)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024640)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025728)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026816)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030976)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079616)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080704)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129344)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130432)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131520)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132608)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394816)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395904)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658112)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659200)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921408)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922496)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184704)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185792)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448000)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449088)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450176)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452288)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976640)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993088)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994176)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995264)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996352)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997440)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998528)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260736)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261824)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262912)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267072)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315712)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316800)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365440)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366528)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367616)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368704)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369792)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373952)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422592)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423680)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472320)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473408)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474496)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475584)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737792)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738880)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001088)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002176)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264384)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265472)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527680)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528768)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790976)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792064)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793152)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795264)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319616)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336064)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337152)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338240)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339328)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340416)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341504)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603712)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604800)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605888)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610048)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658688)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659776)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708416)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709504)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710592)))]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710720)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711808)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236160)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237248)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499456)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500544)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762752)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763840)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026048)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027136)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289344)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290432)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552640)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553728)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554816)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555904)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818112)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821248)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607744)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608832)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609920)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618176)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715392)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716480)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813696)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814784)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815872)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816960)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079168)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080256)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342464)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343552)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605760)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606848)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869056)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870144)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132352)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133440)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134528)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135616)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397824)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400960)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187456)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188544)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189632)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197888)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295104)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296192)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393408)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394496)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 4, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 4, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 4, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([4, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 4, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 4, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; + tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; + tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; + tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; + tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; + tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; + tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; + tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; + tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; + tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; + tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; + tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; + tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; + tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 3, 256])]; + tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, false, true])]; + tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; + tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; + tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; + tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; + tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; + tensor var_245_begin_0 = const()[name = tensor("op_245_begin_0"), val = tensor([0, 3, 0])]; + tensor var_245_end_0 = const()[name = tensor("op_245_end_0"), val = tensor([1, 1, 256])]; + tensor var_245_end_mask_0 = const()[name = tensor("op_245_end_mask_0"), val = tensor([true, true, true])]; + tensor var_245 = slice_by_index(begin = var_245_begin_0, end = var_245_end_0, end_mask = var_245_end_mask_0, x = x_3)[name = tensor("op_245")]; + tensor var_248_begin_0 = const()[name = tensor("op_248_begin_0"), val = tensor([0, 1, 0])]; + tensor var_248_end_0 = const()[name = tensor("op_248_end_0"), val = tensor([1, 16, 256])]; + tensor var_248_end_mask_0 = const()[name = tensor("op_248_end_mask_0"), val = tensor([true, true, true])]; + tensor var_248 = slice_by_index(begin = var_248_begin_0, end = var_248_end_0, end_mask = var_248_end_mask_0, x = window_7)[name = tensor("op_248")]; + tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; + tensor window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_248, var_245))[name = tensor("window_9")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9))[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_273_split_sizes_0 = const()[name = tensor("op_273_split_sizes_0"), val = tensor([256, 256])]; + tensor var_273_axis_0 = const()[name = tensor("op_273_axis_0"), val = tensor(1)]; + tensor var_273_0, tensor var_273_1 = split(axis = var_273_axis_0, split_sizes = var_273_split_sizes_0, x = inputs_3)[name = tensor("op_273")]; + tensor var_275 = sigmoid(x = var_273_1)[name = tensor("op_275")]; + tensor inputs_5 = mul(x = var_273_0, y = var_275)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([4, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([0, -1, 0])]; + tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([4, 16, 256])]; + tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_306 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, x = conv_out_1)[name = tensor("op_306")]; + tensor var_308_perm_0 = const()[name = tensor("op_308_perm_0"), val = tensor([1, 0, 2])]; + tensor var_308 = transpose(perm = var_308_perm_0, x = var_306)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_308)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_331 = const()[name = tensor("op_331"), val = tensor(0x1p-1)]; + tensor var_332 = mul(x = input_39, y = var_331)[name = tensor("op_332")]; + tensor input_41 = add(x = var_332, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor(0x1p-1)]; + tensor var_362 = mul(x = input_51, y = var_361)[name = tensor("op_362")]; + tensor input_53 = add(x = var_362, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_376 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 4, 4, 64])]; + tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_382 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_383 = const()[name = tensor("op_383"), val = tensor(0x1p-3)]; + tensor var_384 = mul(x = var_382, y = var_383)[name = tensor("op_384")]; + tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 4, 4, 64])]; + tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_390 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 4, 4, 64])]; + tensor var_392 = reshape(shape = var_391, x = var_390)[name = tensor("op_392")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_386)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_378)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_402 = const()[name = tensor("op_402"), val = tensor([4, 1])]; + tensor var_403 = reshape(shape = var_402, x = sqrt_s_t_3)[name = tensor("op_403")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_403)[name = tensor("M_3")]; + tensor var_405 = mul(x = qk_3, y = M_3)[name = tensor("op_405")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_392)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_405, y = v_3)[name = tensor("inner_3")]; + tensor var_407_transpose_x_0 = const()[name = tensor("op_407_transpose_x_0"), val = tensor(false)]; + tensor var_407_transpose_y_0 = const()[name = tensor("op_407_transpose_y_0"), val = tensor(false)]; + tensor var_407 = matmul(transpose_x = var_407_transpose_x_0, transpose_y = var_407_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_407")]; + tensor var_408 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_408")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1, 4, 1])]; + tensor var_410 = reshape(shape = var_409, x = var_408)[name = tensor("op_410")]; + tensor cross_3 = mul(x = var_407, y = var_410)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_413 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_413")]; + tensor var_415_transpose_x_1 = const()[name = tensor("op_415_transpose_x_1"), val = tensor(true)]; + tensor var_415_transpose_y_1 = const()[name = tensor("op_415_transpose_y_1"), val = tensor(false)]; + tensor var_415 = matmul(transpose_x = var_415_transpose_x_1, transpose_y = var_415_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_415")]; + tensor new_kv_unnorm_3 = add(x = var_413, y = var_415)[name = tensor("new_kv_unnorm_3")]; + tensor var_417 = const()[name = tensor("op_417"), val = tensor(0x1p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_417)[name = tensor("new_scale_3")]; + tensor var_419 = sqrt(x = new_scale_3)[name = tensor("op_419")]; + tensor var_420 = real_div(x = new_kv_unnorm_3, y = var_419)[name = tensor("op_420")]; + tensor var_421_perm_0 = const()[name = tensor("op_421_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_421 = transpose(perm = var_421_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_421)[name = tensor("out_9")]; + tensor var_425 = const()[name = tensor("op_425"), val = tensor([1, 4, 256])]; + tensor out_11 = reshape(shape = var_425, x = out_9)[name = tensor("out_11")]; + tensor var_427 = silu(x = input_57)[name = tensor("op_427")]; + tensor input_59 = mul(x = var_427, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_11_begin_0 = const()[name = tensor("window_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_11_end_0 = const()[name = tensor("window_11_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_11_end_mask_0 = const()[name = tensor("window_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_11_squeeze_mask_0 = const()[name = tensor("window_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_11 = slice_by_index(begin = window_11_begin_0, end = window_11_end_0, end_mask = window_11_end_mask_0, squeeze_mask = window_11_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_11")]; + tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 0, 0])]; + tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 1, 256])]; + tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; + tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; + tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; + tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; + tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; + tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; + tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; + tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 1, 0])]; + tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 2, 256])]; + tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, false, true])]; + tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; + tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; + tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; + tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; + tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; + tensor var_451_begin_0 = const()[name = tensor("op_451_begin_0"), val = tensor([0, 2, 0])]; + tensor var_451_end_0 = const()[name = tensor("op_451_end_0"), val = tensor([1, 3, 256])]; + tensor var_451_end_mask_0 = const()[name = tensor("op_451_end_mask_0"), val = tensor([true, false, true])]; + tensor var_451 = slice_by_index(begin = var_451_begin_0, end = var_451_end_0, end_mask = var_451_end_mask_0, x = x_9)[name = tensor("op_451")]; + tensor var_454_begin_0 = const()[name = tensor("op_454_begin_0"), val = tensor([0, 1, 0])]; + tensor var_454_end_0 = const()[name = tensor("op_454_end_0"), val = tensor([1, 16, 256])]; + tensor var_454_end_mask_0 = const()[name = tensor("op_454_end_mask_0"), val = tensor([true, true, true])]; + tensor var_454 = slice_by_index(begin = var_454_begin_0, end = var_454_end_0, end_mask = var_454_end_mask_0, x = window_15)[name = tensor("op_454")]; + tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; + tensor window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_454, var_451))[name = tensor("window_17")]; + tensor var_459_begin_0 = const()[name = tensor("op_459_begin_0"), val = tensor([0, 3, 0])]; + tensor var_459_end_0 = const()[name = tensor("op_459_end_0"), val = tensor([1, 1, 256])]; + tensor var_459_end_mask_0 = const()[name = tensor("op_459_end_mask_0"), val = tensor([true, true, true])]; + tensor var_459 = slice_by_index(begin = var_459_begin_0, end = var_459_end_0, end_mask = var_459_end_mask_0, x = x_9)[name = tensor("op_459")]; + tensor var_462_begin_0 = const()[name = tensor("op_462_begin_0"), val = tensor([0, 1, 0])]; + tensor var_462_end_0 = const()[name = tensor("op_462_end_0"), val = tensor([1, 16, 256])]; + tensor var_462_end_mask_0 = const()[name = tensor("op_462_end_mask_0"), val = tensor([true, true, true])]; + tensor var_462 = slice_by_index(begin = var_462_begin_0, end = var_462_end_0, end_mask = var_462_end_mask_0, x = window_17)[name = tensor("op_462")]; + tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; + tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_462, var_459))[name = tensor("window_19")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_13, window_15, window_17, window_19))[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_487_split_sizes_0 = const()[name = tensor("op_487_split_sizes_0"), val = tensor([256, 256])]; + tensor var_487_axis_0 = const()[name = tensor("op_487_axis_0"), val = tensor(1)]; + tensor var_487_0, tensor var_487_1 = split(axis = var_487_axis_0, split_sizes = var_487_split_sizes_0, x = inputs_13)[name = tensor("op_487")]; + tensor var_489 = sigmoid(x = var_487_1)[name = tensor("op_489")]; + tensor inputs_15 = mul(x = var_487_0, y = var_489)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([4, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_520_begin_0 = const()[name = tensor("op_520_begin_0"), val = tensor([0, -1, 0])]; + tensor var_520_end_0 = const()[name = tensor("op_520_end_0"), val = tensor([4, 16, 256])]; + tensor var_520_end_mask_0 = const()[name = tensor("op_520_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_520 = slice_by_index(begin = var_520_begin_0, end = var_520_end_0, end_mask = var_520_end_mask_0, x = conv_out_3)[name = tensor("op_520")]; + tensor var_522_perm_0 = const()[name = tensor("op_522_perm_0"), val = tensor([1, 0, 2])]; + tensor var_522 = transpose(perm = var_522_perm_0, x = var_520)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_522)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_545 = const()[name = tensor("op_545"), val = tensor(0x1p-1)]; + tensor var_546 = mul(x = input_79, y = var_545)[name = tensor("op_546")]; + tensor input_81 = add(x = var_546, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-1)]; + tensor var_576 = mul(x = input_91, y = var_575)[name = tensor("op_576")]; + tensor input_93 = add(x = var_576, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_590 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 4, 4, 64])]; + tensor var_592 = reshape(shape = var_591, x = var_590)[name = tensor("op_592")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_596 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_597 = const()[name = tensor("op_597"), val = tensor(0x1p-3)]; + tensor var_598 = mul(x = var_596, y = var_597)[name = tensor("op_598")]; + tensor var_599 = const()[name = tensor("op_599"), val = tensor([1, 4, 4, 64])]; + tensor var_600 = reshape(shape = var_599, x = var_598)[name = tensor("op_600")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_604 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_605 = const()[name = tensor("op_605"), val = tensor([1, 4, 4, 64])]; + tensor var_606 = reshape(shape = var_605, x = var_604)[name = tensor("op_606")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_600)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_592)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_616 = const()[name = tensor("op_616"), val = tensor([4, 1])]; + tensor var_617 = reshape(shape = var_616, x = sqrt_s_t_5)[name = tensor("op_617")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_617)[name = tensor("M_5")]; + tensor var_619 = mul(x = qk_5, y = M_5)[name = tensor("op_619")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_606)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_619, y = v_5)[name = tensor("inner_5")]; + tensor var_621_transpose_x_0 = const()[name = tensor("op_621_transpose_x_0"), val = tensor(false)]; + tensor var_621_transpose_y_0 = const()[name = tensor("op_621_transpose_y_0"), val = tensor(false)]; + tensor var_621 = matmul(transpose_x = var_621_transpose_x_0, transpose_y = var_621_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_621")]; + tensor var_622 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_622")]; + tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 1, 4, 1])]; + tensor var_624 = reshape(shape = var_623, x = var_622)[name = tensor("op_624")]; + tensor cross_5 = mul(x = var_621, y = var_624)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_627 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_627")]; + tensor var_629_transpose_x_1 = const()[name = tensor("op_629_transpose_x_1"), val = tensor(true)]; + tensor var_629_transpose_y_1 = const()[name = tensor("op_629_transpose_y_1"), val = tensor(false)]; + tensor var_629 = matmul(transpose_x = var_629_transpose_x_1, transpose_y = var_629_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_629")]; + tensor new_kv_unnorm_5 = add(x = var_627, y = var_629)[name = tensor("new_kv_unnorm_5")]; + tensor var_631 = const()[name = tensor("op_631"), val = tensor(0x1p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_631)[name = tensor("new_scale_5")]; + tensor var_633 = sqrt(x = new_scale_5)[name = tensor("op_633")]; + tensor var_634 = real_div(x = new_kv_unnorm_5, y = var_633)[name = tensor("op_634")]; + tensor var_635_perm_0 = const()[name = tensor("op_635_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_635 = transpose(perm = var_635_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_635)[name = tensor("out_15")]; + tensor var_639 = const()[name = tensor("op_639"), val = tensor([1, 4, 256])]; + tensor out_17 = reshape(shape = var_639, x = out_15)[name = tensor("out_17")]; + tensor var_641 = silu(x = input_97)[name = tensor("op_641")]; + tensor input_99 = mul(x = var_641, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_21_begin_0 = const()[name = tensor("window_21_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_21_end_0 = const()[name = tensor("window_21_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_21_end_mask_0 = const()[name = tensor("window_21_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_21_squeeze_mask_0 = const()[name = tensor("window_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_21 = slice_by_index(begin = window_21_begin_0, end = window_21_end_0, end_mask = window_21_end_mask_0, squeeze_mask = window_21_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_21")]; + tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 0, 0])]; + tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; + tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, false, true])]; + tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; + tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; + tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; + tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; + tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; + tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; + tensor var_657_begin_0 = const()[name = tensor("op_657_begin_0"), val = tensor([0, 1, 0])]; + tensor var_657_end_0 = const()[name = tensor("op_657_end_0"), val = tensor([1, 2, 256])]; + tensor var_657_end_mask_0 = const()[name = tensor("op_657_end_mask_0"), val = tensor([true, false, true])]; + tensor var_657 = slice_by_index(begin = var_657_begin_0, end = var_657_end_0, end_mask = var_657_end_mask_0, x = x_15)[name = tensor("op_657")]; + tensor var_660_begin_0 = const()[name = tensor("op_660_begin_0"), val = tensor([0, 1, 0])]; + tensor var_660_end_0 = const()[name = tensor("op_660_end_0"), val = tensor([1, 16, 256])]; + tensor var_660_end_mask_0 = const()[name = tensor("op_660_end_mask_0"), val = tensor([true, true, true])]; + tensor var_660 = slice_by_index(begin = var_660_begin_0, end = var_660_end_0, end_mask = var_660_end_mask_0, x = window_23)[name = tensor("op_660")]; + tensor window_25_interleave_0 = const()[name = tensor("window_25_interleave_0"), val = tensor(false)]; + tensor window_25 = concat(axis = var_26, interleave = window_25_interleave_0, values = (var_660, var_657))[name = tensor("window_25")]; + tensor var_665_begin_0 = const()[name = tensor("op_665_begin_0"), val = tensor([0, 2, 0])]; + tensor var_665_end_0 = const()[name = tensor("op_665_end_0"), val = tensor([1, 3, 256])]; + tensor var_665_end_mask_0 = const()[name = tensor("op_665_end_mask_0"), val = tensor([true, false, true])]; + tensor var_665 = slice_by_index(begin = var_665_begin_0, end = var_665_end_0, end_mask = var_665_end_mask_0, x = x_15)[name = tensor("op_665")]; + tensor var_668_begin_0 = const()[name = tensor("op_668_begin_0"), val = tensor([0, 1, 0])]; + tensor var_668_end_0 = const()[name = tensor("op_668_end_0"), val = tensor([1, 16, 256])]; + tensor var_668_end_mask_0 = const()[name = tensor("op_668_end_mask_0"), val = tensor([true, true, true])]; + tensor var_668 = slice_by_index(begin = var_668_begin_0, end = var_668_end_0, end_mask = var_668_end_mask_0, x = window_25)[name = tensor("op_668")]; + tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; + tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_668, var_665))[name = tensor("window_27")]; + tensor var_673_begin_0 = const()[name = tensor("op_673_begin_0"), val = tensor([0, 3, 0])]; + tensor var_673_end_0 = const()[name = tensor("op_673_end_0"), val = tensor([1, 1, 256])]; + tensor var_673_end_mask_0 = const()[name = tensor("op_673_end_mask_0"), val = tensor([true, true, true])]; + tensor var_673 = slice_by_index(begin = var_673_begin_0, end = var_673_end_0, end_mask = var_673_end_mask_0, x = x_15)[name = tensor("op_673")]; + tensor var_676_begin_0 = const()[name = tensor("op_676_begin_0"), val = tensor([0, 1, 0])]; + tensor var_676_end_0 = const()[name = tensor("op_676_end_0"), val = tensor([1, 16, 256])]; + tensor var_676_end_mask_0 = const()[name = tensor("op_676_end_mask_0"), val = tensor([true, true, true])]; + tensor var_676 = slice_by_index(begin = var_676_begin_0, end = var_676_end_0, end_mask = var_676_end_mask_0, x = window_27)[name = tensor("op_676")]; + tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; + tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_676, var_673))[name = tensor("window_29")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_23, window_25, window_27, window_29))[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_701_split_sizes_0 = const()[name = tensor("op_701_split_sizes_0"), val = tensor([256, 256])]; + tensor var_701_axis_0 = const()[name = tensor("op_701_axis_0"), val = tensor(1)]; + tensor var_701_0, tensor var_701_1 = split(axis = var_701_axis_0, split_sizes = var_701_split_sizes_0, x = inputs_23)[name = tensor("op_701")]; + tensor var_703 = sigmoid(x = var_701_1)[name = tensor("op_703")]; + tensor inputs_25 = mul(x = var_701_0, y = var_703)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([4, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_734_begin_0 = const()[name = tensor("op_734_begin_0"), val = tensor([0, -1, 0])]; + tensor var_734_end_0 = const()[name = tensor("op_734_end_0"), val = tensor([4, 16, 256])]; + tensor var_734_end_mask_0 = const()[name = tensor("op_734_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_734 = slice_by_index(begin = var_734_begin_0, end = var_734_end_0, end_mask = var_734_end_mask_0, x = conv_out_5)[name = tensor("op_734")]; + tensor var_736_perm_0 = const()[name = tensor("op_736_perm_0"), val = tensor([1, 0, 2])]; + tensor var_736 = transpose(perm = var_736_perm_0, x = var_734)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_736)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_759 = const()[name = tensor("op_759"), val = tensor(0x1p-1)]; + tensor var_760 = mul(x = input_119, y = var_759)[name = tensor("op_760")]; + tensor input_121 = add(x = var_760, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_789 = const()[name = tensor("op_789"), val = tensor(0x1p-1)]; + tensor var_790 = mul(x = input_131, y = var_789)[name = tensor("op_790")]; + tensor input_133 = add(x = var_790, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_804 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 4, 4, 64])]; + tensor var_806 = reshape(shape = var_805, x = var_804)[name = tensor("op_806")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_810 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_811 = const()[name = tensor("op_811"), val = tensor(0x1p-3)]; + tensor var_812 = mul(x = var_810, y = var_811)[name = tensor("op_812")]; + tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 4, 4, 64])]; + tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_818 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_819 = const()[name = tensor("op_819"), val = tensor([1, 4, 4, 64])]; + tensor var_820 = reshape(shape = var_819, x = var_818)[name = tensor("op_820")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_814)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_806)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_830 = const()[name = tensor("op_830"), val = tensor([4, 1])]; + tensor var_831 = reshape(shape = var_830, x = sqrt_s_t_7)[name = tensor("op_831")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_831)[name = tensor("M_7")]; + tensor var_833 = mul(x = qk_7, y = M_7)[name = tensor("op_833")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_820)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_833, y = v_7)[name = tensor("inner_7")]; + tensor var_835_transpose_x_0 = const()[name = tensor("op_835_transpose_x_0"), val = tensor(false)]; + tensor var_835_transpose_y_0 = const()[name = tensor("op_835_transpose_y_0"), val = tensor(false)]; + tensor var_835 = matmul(transpose_x = var_835_transpose_x_0, transpose_y = var_835_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_835")]; + tensor var_836 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_836")]; + tensor var_837 = const()[name = tensor("op_837"), val = tensor([1, 1, 4, 1])]; + tensor var_838 = reshape(shape = var_837, x = var_836)[name = tensor("op_838")]; + tensor cross_7 = mul(x = var_835, y = var_838)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_841 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_841")]; + tensor var_843_transpose_x_1 = const()[name = tensor("op_843_transpose_x_1"), val = tensor(true)]; + tensor var_843_transpose_y_1 = const()[name = tensor("op_843_transpose_y_1"), val = tensor(false)]; + tensor var_843 = matmul(transpose_x = var_843_transpose_x_1, transpose_y = var_843_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_843")]; + tensor new_kv_unnorm_7 = add(x = var_841, y = var_843)[name = tensor("new_kv_unnorm_7")]; + tensor var_845 = const()[name = tensor("op_845"), val = tensor(0x1p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_845)[name = tensor("new_scale_7")]; + tensor var_847 = sqrt(x = new_scale_7)[name = tensor("op_847")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_847)[name = tensor("nkv_1")]; + tensor var_849_perm_0 = const()[name = tensor("op_849_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_849 = transpose(perm = var_849_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_849)[name = tensor("out_21")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 4, 256])]; + tensor out_23 = reshape(shape = var_853, x = out_21)[name = tensor("out_23")]; + tensor var_855 = silu(x = input_137)[name = tensor("op_855")]; + tensor input_139 = mul(x = var_855, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_31_begin_0 = const()[name = tensor("window_31_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_31_end_0 = const()[name = tensor("window_31_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_31_end_mask_0 = const()[name = tensor("window_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_31_squeeze_mask_0 = const()[name = tensor("window_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_31 = slice_by_index(begin = window_31_begin_0, end = window_31_end_0, end_mask = window_31_end_mask_0, squeeze_mask = window_31_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_31")]; + tensor var_863_begin_0 = const()[name = tensor("op_863_begin_0"), val = tensor([0, 0, 0])]; + tensor var_863_end_0 = const()[name = tensor("op_863_end_0"), val = tensor([1, 1, 256])]; + tensor var_863_end_mask_0 = const()[name = tensor("op_863_end_mask_0"), val = tensor([true, false, true])]; + tensor var_863 = slice_by_index(begin = var_863_begin_0, end = var_863_end_0, end_mask = var_863_end_mask_0, x = x_21)[name = tensor("op_863")]; + tensor var_866_begin_0 = const()[name = tensor("op_866_begin_0"), val = tensor([0, 1, 0])]; + tensor var_866_end_0 = const()[name = tensor("op_866_end_0"), val = tensor([1, 16, 256])]; + tensor var_866_end_mask_0 = const()[name = tensor("op_866_end_mask_0"), val = tensor([true, true, true])]; + tensor var_866 = slice_by_index(begin = var_866_begin_0, end = var_866_end_0, end_mask = var_866_end_mask_0, x = window_31)[name = tensor("op_866")]; + tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; + tensor window_33 = concat(axis = var_26, interleave = window_33_interleave_0, values = (var_866, var_863))[name = tensor("window_33")]; + tensor var_871_begin_0 = const()[name = tensor("op_871_begin_0"), val = tensor([0, 1, 0])]; + tensor var_871_end_0 = const()[name = tensor("op_871_end_0"), val = tensor([1, 2, 256])]; + tensor var_871_end_mask_0 = const()[name = tensor("op_871_end_mask_0"), val = tensor([true, false, true])]; + tensor var_871 = slice_by_index(begin = var_871_begin_0, end = var_871_end_0, end_mask = var_871_end_mask_0, x = x_21)[name = tensor("op_871")]; + tensor var_874_begin_0 = const()[name = tensor("op_874_begin_0"), val = tensor([0, 1, 0])]; + tensor var_874_end_0 = const()[name = tensor("op_874_end_0"), val = tensor([1, 16, 256])]; + tensor var_874_end_mask_0 = const()[name = tensor("op_874_end_mask_0"), val = tensor([true, true, true])]; + tensor var_874 = slice_by_index(begin = var_874_begin_0, end = var_874_end_0, end_mask = var_874_end_mask_0, x = window_33)[name = tensor("op_874")]; + tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; + tensor window_35 = concat(axis = var_26, interleave = window_35_interleave_0, values = (var_874, var_871))[name = tensor("window_35")]; + tensor var_879_begin_0 = const()[name = tensor("op_879_begin_0"), val = tensor([0, 2, 0])]; + tensor var_879_end_0 = const()[name = tensor("op_879_end_0"), val = tensor([1, 3, 256])]; + tensor var_879_end_mask_0 = const()[name = tensor("op_879_end_mask_0"), val = tensor([true, false, true])]; + tensor var_879 = slice_by_index(begin = var_879_begin_0, end = var_879_end_0, end_mask = var_879_end_mask_0, x = x_21)[name = tensor("op_879")]; + tensor var_882_begin_0 = const()[name = tensor("op_882_begin_0"), val = tensor([0, 1, 0])]; + tensor var_882_end_0 = const()[name = tensor("op_882_end_0"), val = tensor([1, 16, 256])]; + tensor var_882_end_mask_0 = const()[name = tensor("op_882_end_mask_0"), val = tensor([true, true, true])]; + tensor var_882 = slice_by_index(begin = var_882_begin_0, end = var_882_end_0, end_mask = var_882_end_mask_0, x = window_35)[name = tensor("op_882")]; + tensor window_37_interleave_0 = const()[name = tensor("window_37_interleave_0"), val = tensor(false)]; + tensor window_37 = concat(axis = var_26, interleave = window_37_interleave_0, values = (var_882, var_879))[name = tensor("window_37")]; + tensor var_887_begin_0 = const()[name = tensor("op_887_begin_0"), val = tensor([0, 3, 0])]; + tensor var_887_end_0 = const()[name = tensor("op_887_end_0"), val = tensor([1, 1, 256])]; + tensor var_887_end_mask_0 = const()[name = tensor("op_887_end_mask_0"), val = tensor([true, true, true])]; + tensor var_887 = slice_by_index(begin = var_887_begin_0, end = var_887_end_0, end_mask = var_887_end_mask_0, x = x_21)[name = tensor("op_887")]; + tensor var_890_begin_0 = const()[name = tensor("op_890_begin_0"), val = tensor([0, 1, 0])]; + tensor var_890_end_0 = const()[name = tensor("op_890_end_0"), val = tensor([1, 16, 256])]; + tensor var_890_end_mask_0 = const()[name = tensor("op_890_end_mask_0"), val = tensor([true, true, true])]; + tensor var_890 = slice_by_index(begin = var_890_begin_0, end = var_890_end_0, end_mask = var_890_end_mask_0, x = window_37)[name = tensor("op_890")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_890, var_887))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_33, window_35, window_37, window))[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_915_split_sizes_0 = const()[name = tensor("op_915_split_sizes_0"), val = tensor([256, 256])]; + tensor var_915_axis_0 = const()[name = tensor("op_915_axis_0"), val = tensor(1)]; + tensor var_915_0, tensor var_915_1 = split(axis = var_915_axis_0, split_sizes = var_915_split_sizes_0, x = inputs_33)[name = tensor("op_915")]; + tensor var_917 = sigmoid(x = var_915_1)[name = tensor("op_917")]; + tensor inputs_35 = mul(x = var_915_0, y = var_917)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([4, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_948_begin_0 = const()[name = tensor("op_948_begin_0"), val = tensor([0, -1, 0])]; + tensor var_948_end_0 = const()[name = tensor("op_948_end_0"), val = tensor([4, 16, 256])]; + tensor var_948_end_mask_0 = const()[name = tensor("op_948_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_948 = slice_by_index(begin = var_948_begin_0, end = var_948_end_0, end_mask = var_948_end_mask_0, x = conv_out_7)[name = tensor("op_948")]; + tensor var_950_perm_0 = const()[name = tensor("op_950_perm_0"), val = tensor([1, 0, 2])]; + tensor var_950 = transpose(perm = var_950_perm_0, x = var_948)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_950)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_973 = const()[name = tensor("op_973"), val = tensor(0x1p-1)]; + tensor var_974 = mul(x = input_159, y = var_973)[name = tensor("op_974")]; + tensor input_161 = add(x = var_974, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_992_begin_0 = const()[name = tensor("op_992_begin_0"), val = tensor([0, 0, 4])]; + tensor var_992_end_0 = const()[name = tensor("op_992_end_0"), val = tensor([1, 256, 22])]; + tensor var_992_end_mask_0 = const()[name = tensor("op_992_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_992_begin_0, end = var_992_end_0, end_mask = var_992_end_mask_0, x = cat)[name = tensor("op_992")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_994 = const()[name = tensor("op_994"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_995 = reduce_l2_norm(axes = var_994, keep_dims = var_29, x = input_163)[name = tensor("op_995")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_995)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_999_axis_0 = const()[name = tensor("op_999_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_999_axis_0, values = (var_206, var_420, var_634, nkv_1))[name = tensor("op_999")]; + tensor var_1001_axis_0 = const()[name = tensor("op_1001_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1001_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1001")]; + tensor var_1003_axis_0 = const()[name = tensor("op_1003_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1003_axis_0, values = (window_9, window_19, window_29, window))[name = tensor("op_1003")]; + tensor var_1012 = const()[name = tensor("op_1012"), val = tensor(0x1.5798eep-27)]; + tensor var_1017 = const()[name = tensor("op_1017"), val = tensor(0x1.4f8b58p-17)]; + tensor var_1019 = const()[name = tensor("op_1019"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_1020 = const()[name = tensor("op_1020"), val = tensor(true)]; + tensor var_1022 = const()[name = tensor("op_1022"), val = tensor(0x1p+0)]; + tensor var_1026 = const()[name = tensor("op_1026"), val = tensor(-1)]; + tensor var_1032 = const()[name = tensor("op_1032"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395584)))]; + tensor var_1094_axes_0 = const()[name = tensor("op_1094_axes_0"), val = tensor([2])]; + tensor var_1094 = expand_dims(axes = var_1094_axes_0, x = emb)[name = tensor("op_1094")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1094)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_1026, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1102_perm_0 = const()[name = tensor("op_1102_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([12, 4, 256])]; + tensor var_1102 = transpose(perm = var_1102_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1106, x = var_1102)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1114 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1115 = const()[name = tensor("op_1115"), val = tensor([12, 4, 4, 64])]; + tensor var_1116 = reshape(shape = var_1115, x = var_1114)[name = tensor("op_1116")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1120 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1121 = const()[name = tensor("op_1121"), val = tensor(0x1p-3)]; + tensor var_1122 = mul(x = var_1120, y = var_1121)[name = tensor("op_1122")]; + tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([12, 4, 4, 64])]; + tensor var_1124 = reshape(shape = var_1123, x = var_1122)[name = tensor("op_1124")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1128 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1129 = const()[name = tensor("op_1129"), val = tensor([12, 4, 4, 64])]; + tensor var_1130 = reshape(shape = var_1129, x = var_1128)[name = tensor("op_1130")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_1032, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_1022, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1124)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1116)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([1, 4])]; + tensor var_1143 = reshape(shape = var_1142, x = valid_mask)[name = tensor("op_1143")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1143)[name = tensor("causal_with_valid_1")]; + tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([4, 1])]; + tensor var_1146 = reshape(shape = var_1145, x = sqrt_s_t_9)[name = tensor("op_1146")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1146)[name = tensor("M_9")]; + tensor var_1148 = mul(x = qk_9, y = M_9)[name = tensor("op_1148")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1130)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1148, y = v_9)[name = tensor("inner_9")]; + tensor var_1150_transpose_x_0 = const()[name = tensor("op_1150_transpose_x_0"), val = tensor(false)]; + tensor var_1150_transpose_y_0 = const()[name = tensor("op_1150_transpose_y_0"), val = tensor(false)]; + tensor var_1150 = matmul(transpose_x = var_1150_transpose_x_0, transpose_y = var_1150_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1150")]; + tensor var_1151 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1151")]; + tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 1, 4, 1])]; + tensor var_1153 = reshape(shape = var_1152, x = var_1151)[name = tensor("op_1153")]; + tensor cross_9 = mul(x = var_1150, y = var_1153)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 1, 4, 1])]; + tensor var_1157 = reshape(shape = var_1156, x = valid_mask)[name = tensor("op_1157")]; + tensor v_masked_1 = mul(x = v_9, y = var_1157)[name = tensor("v_masked_1")]; + tensor var_1159 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1159")]; + tensor var_1161_transpose_x_1 = const()[name = tensor("op_1161_transpose_x_1"), val = tensor(true)]; + tensor var_1161_transpose_y_1 = const()[name = tensor("op_1161_transpose_y_1"), val = tensor(false)]; + tensor var_1161 = matmul(transpose_x = var_1161_transpose_x_1, transpose_y = var_1161_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1161")]; + tensor new_kv_unnorm_9 = add(x = var_1159, y = var_1161)[name = tensor("new_kv_unnorm_9")]; + tensor var_1163_keep_dims_0 = const()[name = tensor("op_1163_keep_dims_0"), val = tensor(false)]; + tensor var_1163 = reduce_sum(keep_dims = var_1163_keep_dims_0, x = valid_mask)[name = tensor("op_1163")]; + tensor var_1164 = const()[name = tensor("op_1164"), val = tensor([1])]; + tensor var_1165 = reshape(shape = var_1164, x = var_1163)[name = tensor("op_1165")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1165)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_1022, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1169 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1169")]; + tensor var_1170_perm_0 = const()[name = tensor("op_1170_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1170 = transpose(perm = var_1170_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1019, x = var_1170)[name = tensor("out_27")]; + tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([12, 4, 256])]; + tensor out_29 = reshape(shape = var_1174, x = out_27)[name = tensor("out_29")]; + tensor var_1176 = silu(x = input_169)[name = tensor("op_1176")]; + tensor input_171 = mul(x = var_1176, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1017, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1186 = const()[name = tensor("op_1186"), val = tensor([1, 12, 4, 256])]; + tensor var_1187 = reshape(shape = var_1186, x = xt_1)[name = tensor("op_1187")]; + tensor var_1188_perm_0 = const()[name = tensor("op_1188_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([4, 12, 256])]; + tensor var_1188 = transpose(perm = var_1188_perm_0, x = var_1187)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1191, x = var_1188)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1214 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 4, 3, 256])]; + tensor var_1216 = reshape(shape = concat_1, x = var_1214)[name = tensor("op_1216")]; + tensor var_1217_axes_0 = const()[name = tensor("op_1217_axes_0"), val = tensor([0])]; + tensor var_1217 = expand_dims(axes = var_1217_axes_0, x = var_1216)[name = tensor("op_1217")]; + tensor var_1218_perm_0 = const()[name = tensor("op_1218_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1219_axes_0 = const()[name = tensor("op_1219_axes_0"), val = tensor([-2])]; + tensor var_1218 = transpose(perm = var_1218_perm_0, x = var_1217)[name = tensor("transpose_21")]; + tensor var_1219 = squeeze(axes = var_1219_axes_0, x = var_1218)[name = tensor("op_1219")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 4, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1219)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 4, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1219)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 4, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1219)[name = tensor("v_11")]; + tensor var_1227 = const()[name = tensor("op_1227"), val = tensor([12, 16, 64])]; + tensor var_1228 = reshape(shape = var_1227, x = q_11)[name = tensor("op_1228")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1234 = const()[name = tensor("op_1234"), val = tensor([12, 16, 64])]; + tensor var_1235 = reshape(shape = var_1234, x = k_11)[name = tensor("op_1235")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1241 = const()[name = tensor("op_1241"), val = tensor([12, 16, 64])]; + tensor var_1242 = reshape(shape = var_1241, x = v_11)[name = tensor("op_1242")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([4, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1228)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1245, x = q_13)[name = tensor("q_15")]; + tensor var_1247 = const()[name = tensor("op_1247"), val = tensor([4, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1235)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1247, x = k_13)[name = tensor("k_15")]; + tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([4, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1242)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1249, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([2, 0, 1, 3])]; + tensor var_1257 = const()[name = tensor("op_1257"), val = tensor([48, 256])]; + tensor var_1253 = transpose(perm = var_1252, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1257, x = var_1253)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([12, 4, 256])]; + tensor attn_output_7 = reshape(shape = var_1261, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1017, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1017, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1281 = const()[name = tensor("op_1281"), val = tensor([1, 4, 12, 256])]; + tensor x_31 = reshape(shape = var_1281, x = xt_3)[name = tensor("x_31")]; + tensor var_1283_perm_0 = const()[name = tensor("op_1283_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1287 = const()[name = tensor("op_1287"), val = tensor([12, 4, 256])]; + tensor var_1283 = transpose(perm = var_1283_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1287, x = var_1283)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1295 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1296 = const()[name = tensor("op_1296"), val = tensor([12, 4, 4, 64])]; + tensor var_1297 = reshape(shape = var_1296, x = var_1295)[name = tensor("op_1297")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1301 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1302 = const()[name = tensor("op_1302"), val = tensor(0x1p-3)]; + tensor var_1303 = mul(x = var_1301, y = var_1302)[name = tensor("op_1303")]; + tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([12, 4, 4, 64])]; + tensor var_1305 = reshape(shape = var_1304, x = var_1303)[name = tensor("op_1305")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1309 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1310 = const()[name = tensor("op_1310"), val = tensor([12, 4, 4, 64])]; + tensor var_1311 = reshape(shape = var_1310, x = var_1309)[name = tensor("op_1311")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_1022, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1305)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1297)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([4, 1])]; + tensor var_1327 = reshape(shape = var_1326, x = sqrt_s_t)[name = tensor("op_1327")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1327)[name = tensor("M")]; + tensor var_1329 = mul(x = qk, y = M)[name = tensor("op_1329")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1311)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1329, y = v_17)[name = tensor("inner")]; + tensor var_1331_transpose_x_0 = const()[name = tensor("op_1331_transpose_x_0"), val = tensor(false)]; + tensor var_1331_transpose_y_0 = const()[name = tensor("op_1331_transpose_y_0"), val = tensor(false)]; + tensor var_1331 = matmul(transpose_x = var_1331_transpose_x_0, transpose_y = var_1331_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1331")]; + tensor var_1332 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1332")]; + tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 1, 4, 1])]; + tensor var_1334 = reshape(shape = var_1333, x = var_1332)[name = tensor("op_1334")]; + tensor cross = mul(x = var_1331, y = var_1334)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1157)[name = tensor("v_masked")]; + tensor var_1340 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1340")]; + tensor var_1342_transpose_x_1 = const()[name = tensor("op_1342_transpose_x_1"), val = tensor(true)]; + tensor var_1342_transpose_y_1 = const()[name = tensor("op_1342_transpose_y_1"), val = tensor(false)]; + tensor var_1342 = matmul(transpose_x = var_1342_transpose_x_1, transpose_y = var_1342_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1342")]; + tensor new_kv_unnorm = add(x = var_1340, y = var_1342)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1165)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_1022, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1351_perm_0 = const()[name = tensor("op_1351_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1351 = transpose(perm = var_1351_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1019, x = var_1351)[name = tensor("out_33")]; + tensor var_1355 = const()[name = tensor("op_1355"), val = tensor([12, 4, 256])]; + tensor out = reshape(shape = var_1355, x = out_33)[name = tensor("out")]; + tensor var_1357 = silu(x = input_187)[name = tensor("op_1357")]; + tensor input_189 = mul(x = var_1357, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1017, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1367 = const()[name = tensor("op_1367"), val = tensor([1, 12, 4, 256])]; + tensor var_1368 = reshape(shape = var_1367, x = xt_5)[name = tensor("op_1368")]; + tensor var_1369_perm_0 = const()[name = tensor("op_1369_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1372 = const()[name = tensor("op_1372"), val = tensor([4, 12, 256])]; + tensor var_1369 = transpose(perm = var_1369_perm_0, x = var_1368)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1372, x = var_1369)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1395 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 4, 3, 256])]; + tensor var_1397 = reshape(shape = concat_2, x = var_1395)[name = tensor("op_1397")]; + tensor var_1398_axes_0 = const()[name = tensor("op_1398_axes_0"), val = tensor([0])]; + tensor var_1398 = expand_dims(axes = var_1398_axes_0, x = var_1397)[name = tensor("op_1398")]; + tensor var_1399_perm_0 = const()[name = tensor("op_1399_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1400_axes_0 = const()[name = tensor("op_1400_axes_0"), val = tensor([-2])]; + tensor var_1399 = transpose(perm = var_1399_perm_0, x = var_1398)[name = tensor("transpose_8")]; + tensor var_1400 = squeeze(axes = var_1400_axes_0, x = var_1399)[name = tensor("op_1400")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 4, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1400)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 4, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1400)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 4, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1400)[name = tensor("v_19")]; + tensor var_1408 = const()[name = tensor("op_1408"), val = tensor([12, 16, 64])]; + tensor var_1409 = reshape(shape = var_1408, x = q_19)[name = tensor("op_1409")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1415 = const()[name = tensor("op_1415"), val = tensor([12, 16, 64])]; + tensor var_1416 = reshape(shape = var_1415, x = k_19)[name = tensor("op_1416")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1422 = const()[name = tensor("op_1422"), val = tensor([12, 16, 64])]; + tensor var_1423 = reshape(shape = var_1422, x = v_19)[name = tensor("op_1423")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1426 = const()[name = tensor("op_1426"), val = tensor([4, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1409)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1426, x = q_21)[name = tensor("q")]; + tensor var_1428 = const()[name = tensor("op_1428"), val = tensor([4, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1416)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1428, x = k_21)[name = tensor("k")]; + tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([4, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1423)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1430, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([2, 0, 1, 3])]; + tensor var_1438 = const()[name = tensor("op_1438"), val = tensor([48, 256])]; + tensor var_1434 = transpose(perm = var_1433, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1438, x = var_1434)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([12, 4, 256])]; + tensor attn_output = reshape(shape = var_1442, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1017, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1017, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1462 = const()[name = tensor("op_1462"), val = tensor([1, 4, 12, 256])]; + tensor input = reshape(shape = var_1462, x = xt)[name = tensor("input")]; + tensor var_1464 = const()[name = tensor("op_1464"), val = tensor([-1])]; + tensor var_1465 = reduce_l2_norm(axes = var_1464, keep_dims = var_1020, x = input)[name = tensor("op_1465")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_1012, beta = const_42, x = var_1465)[name = tensor("clip_5")]; + tensor var_1467 = real_div(x = input, y = clip_5)[name = tensor("op_1467")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([4, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([4, 256, 12])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1467)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 4, 12])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 4, 11])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1471")]; + tensor var_1473_axis_0 = const()[name = tensor("op_1473_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1473_axis_0, values = (var_1169, nkv))[name = tensor("op_1473")]; + tensor var_1475_axis_0 = const()[name = tensor("op_1475_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1475_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1475")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/weights/weight.bin b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..7915efcdddee50497502d9d86c8bc5285e7b5121 --- /dev/null +++ b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d01e40c100cbf2df86a037acd94dc1f86601a694af5b158b663a46722886f170 +size 44444800 diff --git a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..966abc59c27168f5003194361dd833d1665658b1 --- /dev/null +++ b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:000abbd1df11bb0943f0adfab67500eee955b12db8f10bdc9369e7e808b09657 +size 191053 diff --git a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..7915efcdddee50497502d9d86c8bc5285e7b5121 --- /dev/null +++ b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d01e40c100cbf2df86a037acd94dc1f86601a694af5b158b663a46722886f170 +size 44444800 diff --git a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Manifest.json b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..7f043626a70d0c5eccc20cae217fcc5767d6d198 --- /dev/null +++ b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "7E6A9ABB-631C-4011-875E-660D5DCE8EFC": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "A7E5EDC0-82DC-425C-A42A-618F7B2317F7": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "7E6A9ABB-631C-4011-875E-660D5DCE8EFC" +} diff --git a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/analytics/coremldata.bin b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..4d38c0949a58bc566d5e4794910ebd4ae04a7a27 --- /dev/null +++ b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f8e24abfe7f695ffad28e2adc4c3072705f7bf9cf80102cc782f8e933e72f27 +size 243 diff --git a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/coremldata.bin b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..8915b742fe103cc63e86c1d10e5b105b8647f78d --- /dev/null +++ b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9da2e560b5f0570964bbcfe58e5520dfcb2b29af2679608619c52ae68309c02 +size 1310 diff --git a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/metadata.json b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..bbd1303a91af367bee74f4f22f06f98437f409e1 --- /dev/null +++ b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/metadata.json @@ -0,0 +1,218 @@ +[ + { + "metadataOutputVersion" : "3.0", + "shortDescription" : "LS-EEND DIHARD III streaming diarizer (pipeline, T=5, max_speakers=10)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 5 × 10)", + "shortDescription" : "", + "shape" : "[1, 5, 10]", + "name" : "probs", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv_new", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale_new", + "type" : "MultiArray" + } + ], + "storagePrecision" : "Float32", + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios16.reduceL2Norm" : 2, + "Ios17.reshape" : 68, + "Ios16.softmax" : 2, + "Ios17.matmul" : 29, + "Ios17.transpose" : 57, + "Split" : 4, + "Ios17.expandDims" : 3, + "Ios17.add" : 46, + "Ios16.sigmoid" : 5, + "Ios17.sliceByIndex" : 72, + "Tile" : 1, + "Ios16.reduceSum" : 1, + "Ios17.squeeze" : 2, + "Ios17.layerNorm" : 33, + "Ios17.batchNorm" : 4, + "Ios17.sqrt" : 18, + "Ios17.conv" : 13, + "Ios17.clip" : 6, + "Ios16.silu" : 18, + "Ios17.realDiv" : 20, + "Ios17.linear" : 56, + "Stack" : 5, + "Ios17.concat" : 26, + "Ios16.relu" : 2, + "Ios16.cumsum" : 1, + "Ios17.mul" : 47 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 5 × 345)", + "shortDescription" : "", + "shape" : "[1, 5, 345]", + "name" : "features", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[4, 1, 4, 64, 64]", + "name" : "enc_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1)", + "shortDescription" : "", + "shape" : "[4, 1]", + "name" : "enc_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 4 × 1 × 16 × 256)", + "shortDescription" : "", + "shape" : "[4, 1, 16, 256]", + "name" : "enc_conv_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 18)", + "shortDescription" : "", + "shape" : "[1, 256, 18]", + "name" : "cnn_window", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 12 × 4 × 64 × 64)", + "shortDescription" : "", + "shape" : "[2, 12, 4, 64, 64]", + "name" : "dec_kv", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1)", + "shortDescription" : "", + "shape" : "[2, 1]", + "name" : "dec_scale", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 5)", + "shortDescription" : "", + "shape" : "[5]", + "name" : "valid_mask", + "type" : "MultiArray" + } + ], + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-16", + "config" : "{\"model_name\": \"dih3\", \"model_label\": \"DIHARD III\", \"variant\": \"pipeline\", \"chunk_size\": 5, \"step_duration_ms\": 500, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.source" : "torch==2.6.0", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "generatedClassName" : "ls_eend_dih3_500ms", + "method" : "predict" + } +] \ No newline at end of file diff --git a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/model.mil b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..d06d8cb5807c4c0b54c3e07f38ba5655c925422a --- /dev/null +++ b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/model.mil @@ -0,0 +1,1366 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2, 0x1.4p+2])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982144)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983232)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336576)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337664)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338752)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339840)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340928)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345088)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393728)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394816)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443456)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444544)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445632)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446720)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708928)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7710016)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972224)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973312)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235520)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236608)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498816)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499904)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762112)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763200)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764288)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766400)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290752)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307200)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308288)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309376)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310464)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311552)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312640)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574848)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575936)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9577024)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581184)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629824)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630912)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679552)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680640)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681728)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682816)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683904)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688064)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736704)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737792)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786432)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787520)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788608)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789696)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051904)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052992)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315200)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316288)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578496)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579584)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841792)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842880)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105088)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106176)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107264)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109376)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633728)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650176)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651264)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652352)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653440)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654528)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655616)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917824)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918912)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15920000)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924160)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972800)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973888)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022528)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023616)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024704)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025792)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026880)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18031040)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079680)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080768)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129408)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130496)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131584)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132672)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394880)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395968)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658176)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659264)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921472)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922560)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184768)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185856)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448064)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449152)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450240)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452352)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976704)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993152)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994240)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995328)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996416)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997504)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998592)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260800)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261888)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262976)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267136)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315776)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316864)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365504)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366592)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367680)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368768)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369856)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24374016)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422656)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423744)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472384)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473472)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474560)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475648)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737856)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738944)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001152)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002240)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264448)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265536)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527744)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528832)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791040)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792128)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793216)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795328)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319680)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336128)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337216)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338304)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339392)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340480)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341568)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603776)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604864)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605952)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610112)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658752)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659840)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708480)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709568)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710656)))]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710848)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711936)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236288)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237376)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499584)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500672)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762880)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763968)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026176)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027264)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289472)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290560)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552768)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553856)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554944)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32556032)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818240)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821376)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607872)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608960)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33610048)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618304)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715520)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716608)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813824)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814912)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816000)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37817088)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079296)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080384)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342592)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343680)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605888)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606976)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869184)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870272)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132480)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133568)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134656)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135744)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397952)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39401088)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187584)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188672)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189760)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40198016)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295232)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296320)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393536)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394624)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_18 = const()[name = tensor("op_18"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_21 = const()[name = tensor("op_21"), val = tensor(2)]; + tensor var_24 = const()[name = tensor("op_24"), val = tensor(0)]; + tensor var_27 = const()[name = tensor("op_27"), val = tensor(1)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(0x1.4f8b58p-17)]; + tensor var_30 = const()[name = tensor("op_30"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_29, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_148 = const()[name = tensor("op_148"), val = tensor(0x1p-1)]; + tensor var_149 = mul(x = input_11, y = var_148)[name = tensor("op_149")]; + tensor input_13 = add(x = var_149, y = input_3)[name = tensor("input_13")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_29, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_163 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_164 = const()[name = tensor("op_164"), val = tensor([1, 5, 4, 64])]; + tensor var_165 = reshape(shape = var_164, x = var_163)[name = tensor("op_165")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_169 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_170 = const()[name = tensor("op_170"), val = tensor(0x1p-3)]; + tensor var_171 = mul(x = var_169, y = var_170)[name = tensor("op_171")]; + tensor var_172 = const()[name = tensor("op_172"), val = tensor([1, 5, 4, 64])]; + tensor var_173 = reshape(shape = var_172, x = var_171)[name = tensor("op_173")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_177 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_178 = const()[name = tensor("op_178"), val = tensor([1, 5, 4, 64])]; + tensor var_179 = reshape(shape = var_178, x = var_177)[name = tensor("op_179")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_173)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_165)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_189 = const()[name = tensor("op_189"), val = tensor([5, 1])]; + tensor var_190 = reshape(shape = var_189, x = sqrt_s_t_1)[name = tensor("op_190")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_190)[name = tensor("M_1")]; + tensor var_192 = mul(x = qk_1, y = M_1)[name = tensor("op_192")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_179)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_192, y = v_1)[name = tensor("inner_1")]; + tensor var_194_transpose_x_0 = const()[name = tensor("op_194_transpose_x_0"), val = tensor(false)]; + tensor var_194_transpose_y_0 = const()[name = tensor("op_194_transpose_y_0"), val = tensor(false)]; + tensor var_194 = matmul(transpose_x = var_194_transpose_x_0, transpose_y = var_194_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_194")]; + tensor var_195 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_195")]; + tensor var_196 = const()[name = tensor("op_196"), val = tensor([1, 1, 5, 1])]; + tensor var_197 = reshape(shape = var_196, x = var_195)[name = tensor("op_197")]; + tensor cross_1 = mul(x = var_194, y = var_197)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_200 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_200")]; + tensor var_202_transpose_x_1 = const()[name = tensor("op_202_transpose_x_1"), val = tensor(true)]; + tensor var_202_transpose_y_1 = const()[name = tensor("op_202_transpose_y_1"), val = tensor(false)]; + tensor var_202 = matmul(transpose_x = var_202_transpose_x_1, transpose_y = var_202_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_202")]; + tensor new_kv_unnorm_1 = add(x = var_200, y = var_202)[name = tensor("new_kv_unnorm_1")]; + tensor var_204 = const()[name = tensor("op_204"), val = tensor(0x1.4p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_204)[name = tensor("new_scale_1")]; + tensor var_206 = sqrt(x = new_scale_1)[name = tensor("op_206")]; + tensor var_207 = real_div(x = new_kv_unnorm_1, y = var_206)[name = tensor("op_207")]; + tensor var_208_perm_0 = const()[name = tensor("op_208_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_208 = transpose(perm = var_208_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_18, x = var_208)[name = tensor("out_3")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor([1, 5, 256])]; + tensor out_5 = reshape(shape = var_212, x = out_3)[name = tensor("out_5")]; + tensor var_214 = silu(x = input_17)[name = tensor("op_214")]; + tensor input_19 = mul(x = var_214, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_222_begin_0 = const()[name = tensor("op_222_begin_0"), val = tensor([0, 0, 0])]; + tensor var_222_end_0 = const()[name = tensor("op_222_end_0"), val = tensor([1, 1, 256])]; + tensor var_222_end_mask_0 = const()[name = tensor("op_222_end_mask_0"), val = tensor([true, false, true])]; + tensor var_222 = slice_by_index(begin = var_222_begin_0, end = var_222_end_0, end_mask = var_222_end_mask_0, x = x_3)[name = tensor("op_222")]; + tensor var_225_begin_0 = const()[name = tensor("op_225_begin_0"), val = tensor([0, 1, 0])]; + tensor var_225_end_0 = const()[name = tensor("op_225_end_0"), val = tensor([1, 16, 256])]; + tensor var_225_end_mask_0 = const()[name = tensor("op_225_end_mask_0"), val = tensor([true, true, true])]; + tensor var_225 = slice_by_index(begin = var_225_begin_0, end = var_225_end_0, end_mask = var_225_end_mask_0, x = window_1)[name = tensor("op_225")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_27, interleave = window_3_interleave_0, values = (var_225, var_222))[name = tensor("window_3")]; + tensor var_230_begin_0 = const()[name = tensor("op_230_begin_0"), val = tensor([0, 1, 0])]; + tensor var_230_end_0 = const()[name = tensor("op_230_end_0"), val = tensor([1, 2, 256])]; + tensor var_230_end_mask_0 = const()[name = tensor("op_230_end_mask_0"), val = tensor([true, false, true])]; + tensor var_230 = slice_by_index(begin = var_230_begin_0, end = var_230_end_0, end_mask = var_230_end_mask_0, x = x_3)[name = tensor("op_230")]; + tensor var_233_begin_0 = const()[name = tensor("op_233_begin_0"), val = tensor([0, 1, 0])]; + tensor var_233_end_0 = const()[name = tensor("op_233_end_0"), val = tensor([1, 16, 256])]; + tensor var_233_end_mask_0 = const()[name = tensor("op_233_end_mask_0"), val = tensor([true, true, true])]; + tensor var_233 = slice_by_index(begin = var_233_begin_0, end = var_233_end_0, end_mask = var_233_end_mask_0, x = window_3)[name = tensor("op_233")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_27, interleave = window_5_interleave_0, values = (var_233, var_230))[name = tensor("window_5")]; + tensor var_238_begin_0 = const()[name = tensor("op_238_begin_0"), val = tensor([0, 2, 0])]; + tensor var_238_end_0 = const()[name = tensor("op_238_end_0"), val = tensor([1, 3, 256])]; + tensor var_238_end_mask_0 = const()[name = tensor("op_238_end_mask_0"), val = tensor([true, false, true])]; + tensor var_238 = slice_by_index(begin = var_238_begin_0, end = var_238_end_0, end_mask = var_238_end_mask_0, x = x_3)[name = tensor("op_238")]; + tensor var_241_begin_0 = const()[name = tensor("op_241_begin_0"), val = tensor([0, 1, 0])]; + tensor var_241_end_0 = const()[name = tensor("op_241_end_0"), val = tensor([1, 16, 256])]; + tensor var_241_end_mask_0 = const()[name = tensor("op_241_end_mask_0"), val = tensor([true, true, true])]; + tensor var_241 = slice_by_index(begin = var_241_begin_0, end = var_241_end_0, end_mask = var_241_end_mask_0, x = window_5)[name = tensor("op_241")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_27, interleave = window_7_interleave_0, values = (var_241, var_238))[name = tensor("window_7")]; + tensor var_246_begin_0 = const()[name = tensor("op_246_begin_0"), val = tensor([0, 3, 0])]; + tensor var_246_end_0 = const()[name = tensor("op_246_end_0"), val = tensor([1, 4, 256])]; + tensor var_246_end_mask_0 = const()[name = tensor("op_246_end_mask_0"), val = tensor([true, false, true])]; + tensor var_246 = slice_by_index(begin = var_246_begin_0, end = var_246_end_0, end_mask = var_246_end_mask_0, x = x_3)[name = tensor("op_246")]; + tensor var_249_begin_0 = const()[name = tensor("op_249_begin_0"), val = tensor([0, 1, 0])]; + tensor var_249_end_0 = const()[name = tensor("op_249_end_0"), val = tensor([1, 16, 256])]; + tensor var_249_end_mask_0 = const()[name = tensor("op_249_end_mask_0"), val = tensor([true, true, true])]; + tensor var_249 = slice_by_index(begin = var_249_begin_0, end = var_249_end_0, end_mask = var_249_end_mask_0, x = window_7)[name = tensor("op_249")]; + tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; + tensor window_9 = concat(axis = var_27, interleave = window_9_interleave_0, values = (var_249, var_246))[name = tensor("window_9")]; + tensor var_254_begin_0 = const()[name = tensor("op_254_begin_0"), val = tensor([0, 4, 0])]; + tensor var_254_end_0 = const()[name = tensor("op_254_end_0"), val = tensor([1, 1, 256])]; + tensor var_254_end_mask_0 = const()[name = tensor("op_254_end_mask_0"), val = tensor([true, true, true])]; + tensor var_254 = slice_by_index(begin = var_254_begin_0, end = var_254_end_0, end_mask = var_254_end_mask_0, x = x_3)[name = tensor("op_254")]; + tensor var_257_begin_0 = const()[name = tensor("op_257_begin_0"), val = tensor([0, 1, 0])]; + tensor var_257_end_0 = const()[name = tensor("op_257_end_0"), val = tensor([1, 16, 256])]; + tensor var_257_end_mask_0 = const()[name = tensor("op_257_end_mask_0"), val = tensor([true, true, true])]; + tensor var_257 = slice_by_index(begin = var_257_begin_0, end = var_257_end_0, end_mask = var_257_end_mask_0, x = window_9)[name = tensor("op_257")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_27, interleave = window_11_interleave_0, values = (var_257, var_254))[name = tensor("window_11")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_24, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9, window_11))[name = tensor("input_21")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_282_split_sizes_0 = const()[name = tensor("op_282_split_sizes_0"), val = tensor([256, 256])]; + tensor var_282_axis_0 = const()[name = tensor("op_282_axis_0"), val = tensor(1)]; + tensor var_282_0, tensor var_282_1 = split(axis = var_282_axis_0, split_sizes = var_282_split_sizes_0, x = inputs_3)[name = tensor("op_282")]; + tensor var_284 = sigmoid(x = var_282_1)[name = tensor("op_284")]; + tensor inputs_5 = mul(x = var_282_0, y = var_284)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([5, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_315_begin_0 = const()[name = tensor("op_315_begin_0"), val = tensor([0, -1, 0])]; + tensor var_315_end_0 = const()[name = tensor("op_315_end_0"), val = tensor([5, 16, 256])]; + tensor var_315_end_mask_0 = const()[name = tensor("op_315_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_315 = slice_by_index(begin = var_315_begin_0, end = var_315_end_0, end_mask = var_315_end_mask_0, x = conv_out_1)[name = tensor("op_315")]; + tensor var_317_perm_0 = const()[name = tensor("op_317_perm_0"), val = tensor([1, 0, 2])]; + tensor var_317 = transpose(perm = var_317_perm_0, x = var_315)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_317)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor(0x1p-1)]; + tensor var_341 = mul(x = input_39, y = var_340)[name = tensor("op_341")]; + tensor input_41 = add(x = var_341, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_29, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_370 = const()[name = tensor("op_370"), val = tensor(0x1p-1)]; + tensor var_371 = mul(x = input_51, y = var_370)[name = tensor("op_371")]; + tensor input_53 = add(x = var_371, y = input_43)[name = tensor("input_53")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_29, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_385 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([1, 5, 4, 64])]; + tensor var_387 = reshape(shape = var_386, x = var_385)[name = tensor("op_387")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_391 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_392 = const()[name = tensor("op_392"), val = tensor(0x1p-3)]; + tensor var_393 = mul(x = var_391, y = var_392)[name = tensor("op_393")]; + tensor var_394 = const()[name = tensor("op_394"), val = tensor([1, 5, 4, 64])]; + tensor var_395 = reshape(shape = var_394, x = var_393)[name = tensor("op_395")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_399 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_400 = const()[name = tensor("op_400"), val = tensor([1, 5, 4, 64])]; + tensor var_401 = reshape(shape = var_400, x = var_399)[name = tensor("op_401")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_395)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_387)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([5, 1])]; + tensor var_412 = reshape(shape = var_411, x = sqrt_s_t_3)[name = tensor("op_412")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_412)[name = tensor("M_3")]; + tensor var_414 = mul(x = qk_3, y = M_3)[name = tensor("op_414")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_401)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_414, y = v_3)[name = tensor("inner_3")]; + tensor var_416_transpose_x_0 = const()[name = tensor("op_416_transpose_x_0"), val = tensor(false)]; + tensor var_416_transpose_y_0 = const()[name = tensor("op_416_transpose_y_0"), val = tensor(false)]; + tensor var_416 = matmul(transpose_x = var_416_transpose_x_0, transpose_y = var_416_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_416")]; + tensor var_417 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_417")]; + tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1, 5, 1])]; + tensor var_419 = reshape(shape = var_418, x = var_417)[name = tensor("op_419")]; + tensor cross_3 = mul(x = var_416, y = var_419)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_422 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_422")]; + tensor var_424_transpose_x_1 = const()[name = tensor("op_424_transpose_x_1"), val = tensor(true)]; + tensor var_424_transpose_y_1 = const()[name = tensor("op_424_transpose_y_1"), val = tensor(false)]; + tensor var_424 = matmul(transpose_x = var_424_transpose_x_1, transpose_y = var_424_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_424")]; + tensor new_kv_unnorm_3 = add(x = var_422, y = var_424)[name = tensor("new_kv_unnorm_3")]; + tensor var_426 = const()[name = tensor("op_426"), val = tensor(0x1.4p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_426)[name = tensor("new_scale_3")]; + tensor var_428 = sqrt(x = new_scale_3)[name = tensor("op_428")]; + tensor var_429 = real_div(x = new_kv_unnorm_3, y = var_428)[name = tensor("op_429")]; + tensor var_430_perm_0 = const()[name = tensor("op_430_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_430 = transpose(perm = var_430_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_18, x = var_430)[name = tensor("out_9")]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 5, 256])]; + tensor out_11 = reshape(shape = var_434, x = out_9)[name = tensor("out_11")]; + tensor var_436 = silu(x = input_57)[name = tensor("op_436")]; + tensor input_59 = mul(x = var_436, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; + tensor var_444_begin_0 = const()[name = tensor("op_444_begin_0"), val = tensor([0, 0, 0])]; + tensor var_444_end_0 = const()[name = tensor("op_444_end_0"), val = tensor([1, 1, 256])]; + tensor var_444_end_mask_0 = const()[name = tensor("op_444_end_mask_0"), val = tensor([true, false, true])]; + tensor var_444 = slice_by_index(begin = var_444_begin_0, end = var_444_end_0, end_mask = var_444_end_mask_0, x = x_9)[name = tensor("op_444")]; + tensor var_447_begin_0 = const()[name = tensor("op_447_begin_0"), val = tensor([0, 1, 0])]; + tensor var_447_end_0 = const()[name = tensor("op_447_end_0"), val = tensor([1, 16, 256])]; + tensor var_447_end_mask_0 = const()[name = tensor("op_447_end_mask_0"), val = tensor([true, true, true])]; + tensor var_447 = slice_by_index(begin = var_447_begin_0, end = var_447_end_0, end_mask = var_447_end_mask_0, x = window_13)[name = tensor("op_447")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_27, interleave = window_15_interleave_0, values = (var_447, var_444))[name = tensor("window_15")]; + tensor var_452_begin_0 = const()[name = tensor("op_452_begin_0"), val = tensor([0, 1, 0])]; + tensor var_452_end_0 = const()[name = tensor("op_452_end_0"), val = tensor([1, 2, 256])]; + tensor var_452_end_mask_0 = const()[name = tensor("op_452_end_mask_0"), val = tensor([true, false, true])]; + tensor var_452 = slice_by_index(begin = var_452_begin_0, end = var_452_end_0, end_mask = var_452_end_mask_0, x = x_9)[name = tensor("op_452")]; + tensor var_455_begin_0 = const()[name = tensor("op_455_begin_0"), val = tensor([0, 1, 0])]; + tensor var_455_end_0 = const()[name = tensor("op_455_end_0"), val = tensor([1, 16, 256])]; + tensor var_455_end_mask_0 = const()[name = tensor("op_455_end_mask_0"), val = tensor([true, true, true])]; + tensor var_455 = slice_by_index(begin = var_455_begin_0, end = var_455_end_0, end_mask = var_455_end_mask_0, x = window_15)[name = tensor("op_455")]; + tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; + tensor window_17 = concat(axis = var_27, interleave = window_17_interleave_0, values = (var_455, var_452))[name = tensor("window_17")]; + tensor var_460_begin_0 = const()[name = tensor("op_460_begin_0"), val = tensor([0, 2, 0])]; + tensor var_460_end_0 = const()[name = tensor("op_460_end_0"), val = tensor([1, 3, 256])]; + tensor var_460_end_mask_0 = const()[name = tensor("op_460_end_mask_0"), val = tensor([true, false, true])]; + tensor var_460 = slice_by_index(begin = var_460_begin_0, end = var_460_end_0, end_mask = var_460_end_mask_0, x = x_9)[name = tensor("op_460")]; + tensor var_463_begin_0 = const()[name = tensor("op_463_begin_0"), val = tensor([0, 1, 0])]; + tensor var_463_end_0 = const()[name = tensor("op_463_end_0"), val = tensor([1, 16, 256])]; + tensor var_463_end_mask_0 = const()[name = tensor("op_463_end_mask_0"), val = tensor([true, true, true])]; + tensor var_463 = slice_by_index(begin = var_463_begin_0, end = var_463_end_0, end_mask = var_463_end_mask_0, x = window_17)[name = tensor("op_463")]; + tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; + tensor window_19 = concat(axis = var_27, interleave = window_19_interleave_0, values = (var_463, var_460))[name = tensor("window_19")]; + tensor var_468_begin_0 = const()[name = tensor("op_468_begin_0"), val = tensor([0, 3, 0])]; + tensor var_468_end_0 = const()[name = tensor("op_468_end_0"), val = tensor([1, 4, 256])]; + tensor var_468_end_mask_0 = const()[name = tensor("op_468_end_mask_0"), val = tensor([true, false, true])]; + tensor var_468 = slice_by_index(begin = var_468_begin_0, end = var_468_end_0, end_mask = var_468_end_mask_0, x = x_9)[name = tensor("op_468")]; + tensor var_471_begin_0 = const()[name = tensor("op_471_begin_0"), val = tensor([0, 1, 0])]; + tensor var_471_end_0 = const()[name = tensor("op_471_end_0"), val = tensor([1, 16, 256])]; + tensor var_471_end_mask_0 = const()[name = tensor("op_471_end_mask_0"), val = tensor([true, true, true])]; + tensor var_471 = slice_by_index(begin = var_471_begin_0, end = var_471_end_0, end_mask = var_471_end_mask_0, x = window_19)[name = tensor("op_471")]; + tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; + tensor window_21 = concat(axis = var_27, interleave = window_21_interleave_0, values = (var_471, var_468))[name = tensor("window_21")]; + tensor var_476_begin_0 = const()[name = tensor("op_476_begin_0"), val = tensor([0, 4, 0])]; + tensor var_476_end_0 = const()[name = tensor("op_476_end_0"), val = tensor([1, 1, 256])]; + tensor var_476_end_mask_0 = const()[name = tensor("op_476_end_mask_0"), val = tensor([true, true, true])]; + tensor var_476 = slice_by_index(begin = var_476_begin_0, end = var_476_end_0, end_mask = var_476_end_mask_0, x = x_9)[name = tensor("op_476")]; + tensor var_479_begin_0 = const()[name = tensor("op_479_begin_0"), val = tensor([0, 1, 0])]; + tensor var_479_end_0 = const()[name = tensor("op_479_end_0"), val = tensor([1, 16, 256])]; + tensor var_479_end_mask_0 = const()[name = tensor("op_479_end_mask_0"), val = tensor([true, true, true])]; + tensor var_479 = slice_by_index(begin = var_479_begin_0, end = var_479_end_0, end_mask = var_479_end_mask_0, x = window_21)[name = tensor("op_479")]; + tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; + tensor window_23 = concat(axis = var_27, interleave = window_23_interleave_0, values = (var_479, var_476))[name = tensor("window_23")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_24, interleave = input_61_interleave_0, values = (window_15, window_17, window_19, window_21, window_23))[name = tensor("input_61")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_504_split_sizes_0 = const()[name = tensor("op_504_split_sizes_0"), val = tensor([256, 256])]; + tensor var_504_axis_0 = const()[name = tensor("op_504_axis_0"), val = tensor(1)]; + tensor var_504_0, tensor var_504_1 = split(axis = var_504_axis_0, split_sizes = var_504_split_sizes_0, x = inputs_13)[name = tensor("op_504")]; + tensor var_506 = sigmoid(x = var_504_1)[name = tensor("op_506")]; + tensor inputs_15 = mul(x = var_504_0, y = var_506)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([5, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_537_begin_0 = const()[name = tensor("op_537_begin_0"), val = tensor([0, -1, 0])]; + tensor var_537_end_0 = const()[name = tensor("op_537_end_0"), val = tensor([5, 16, 256])]; + tensor var_537_end_mask_0 = const()[name = tensor("op_537_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_537 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = conv_out_3)[name = tensor("op_537")]; + tensor var_539_perm_0 = const()[name = tensor("op_539_perm_0"), val = tensor([1, 0, 2])]; + tensor var_539 = transpose(perm = var_539_perm_0, x = var_537)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_539)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_562 = const()[name = tensor("op_562"), val = tensor(0x1p-1)]; + tensor var_563 = mul(x = input_79, y = var_562)[name = tensor("op_563")]; + tensor input_81 = add(x = var_563, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_29, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_592 = const()[name = tensor("op_592"), val = tensor(0x1p-1)]; + tensor var_593 = mul(x = input_91, y = var_592)[name = tensor("op_593")]; + tensor input_93 = add(x = var_593, y = input_83)[name = tensor("input_93")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_29, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_607 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_608 = const()[name = tensor("op_608"), val = tensor([1, 5, 4, 64])]; + tensor var_609 = reshape(shape = var_608, x = var_607)[name = tensor("op_609")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_613 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_614 = const()[name = tensor("op_614"), val = tensor(0x1p-3)]; + tensor var_615 = mul(x = var_613, y = var_614)[name = tensor("op_615")]; + tensor var_616 = const()[name = tensor("op_616"), val = tensor([1, 5, 4, 64])]; + tensor var_617 = reshape(shape = var_616, x = var_615)[name = tensor("op_617")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_621 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_622 = const()[name = tensor("op_622"), val = tensor([1, 5, 4, 64])]; + tensor var_623 = reshape(shape = var_622, x = var_621)[name = tensor("op_623")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_617)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_609)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_633 = const()[name = tensor("op_633"), val = tensor([5, 1])]; + tensor var_634 = reshape(shape = var_633, x = sqrt_s_t_5)[name = tensor("op_634")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_634)[name = tensor("M_5")]; + tensor var_636 = mul(x = qk_5, y = M_5)[name = tensor("op_636")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_623)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_636, y = v_5)[name = tensor("inner_5")]; + tensor var_638_transpose_x_0 = const()[name = tensor("op_638_transpose_x_0"), val = tensor(false)]; + tensor var_638_transpose_y_0 = const()[name = tensor("op_638_transpose_y_0"), val = tensor(false)]; + tensor var_638 = matmul(transpose_x = var_638_transpose_x_0, transpose_y = var_638_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_638")]; + tensor var_639 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_639")]; + tensor var_640 = const()[name = tensor("op_640"), val = tensor([1, 1, 5, 1])]; + tensor var_641 = reshape(shape = var_640, x = var_639)[name = tensor("op_641")]; + tensor cross_5 = mul(x = var_638, y = var_641)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_644 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_644")]; + tensor var_646_transpose_x_1 = const()[name = tensor("op_646_transpose_x_1"), val = tensor(true)]; + tensor var_646_transpose_y_1 = const()[name = tensor("op_646_transpose_y_1"), val = tensor(false)]; + tensor var_646 = matmul(transpose_x = var_646_transpose_x_1, transpose_y = var_646_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_646")]; + tensor new_kv_unnorm_5 = add(x = var_644, y = var_646)[name = tensor("new_kv_unnorm_5")]; + tensor var_648 = const()[name = tensor("op_648"), val = tensor(0x1.4p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_648)[name = tensor("new_scale_5")]; + tensor var_650 = sqrt(x = new_scale_5)[name = tensor("op_650")]; + tensor var_651 = real_div(x = new_kv_unnorm_5, y = var_650)[name = tensor("op_651")]; + tensor var_652_perm_0 = const()[name = tensor("op_652_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_652 = transpose(perm = var_652_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_18, x = var_652)[name = tensor("out_15")]; + tensor var_656 = const()[name = tensor("op_656"), val = tensor([1, 5, 256])]; + tensor out_17 = reshape(shape = var_656, x = out_15)[name = tensor("out_17")]; + tensor var_658 = silu(x = input_97)[name = tensor("op_658")]; + tensor input_99 = mul(x = var_658, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; + tensor var_666_begin_0 = const()[name = tensor("op_666_begin_0"), val = tensor([0, 0, 0])]; + tensor var_666_end_0 = const()[name = tensor("op_666_end_0"), val = tensor([1, 1, 256])]; + tensor var_666_end_mask_0 = const()[name = tensor("op_666_end_mask_0"), val = tensor([true, false, true])]; + tensor var_666 = slice_by_index(begin = var_666_begin_0, end = var_666_end_0, end_mask = var_666_end_mask_0, x = x_15)[name = tensor("op_666")]; + tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 1, 0])]; + tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 16, 256])]; + tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, true])]; + tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = window_25)[name = tensor("op_669")]; + tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; + tensor window_27 = concat(axis = var_27, interleave = window_27_interleave_0, values = (var_669, var_666))[name = tensor("window_27")]; + tensor var_674_begin_0 = const()[name = tensor("op_674_begin_0"), val = tensor([0, 1, 0])]; + tensor var_674_end_0 = const()[name = tensor("op_674_end_0"), val = tensor([1, 2, 256])]; + tensor var_674_end_mask_0 = const()[name = tensor("op_674_end_mask_0"), val = tensor([true, false, true])]; + tensor var_674 = slice_by_index(begin = var_674_begin_0, end = var_674_end_0, end_mask = var_674_end_mask_0, x = x_15)[name = tensor("op_674")]; + tensor var_677_begin_0 = const()[name = tensor("op_677_begin_0"), val = tensor([0, 1, 0])]; + tensor var_677_end_0 = const()[name = tensor("op_677_end_0"), val = tensor([1, 16, 256])]; + tensor var_677_end_mask_0 = const()[name = tensor("op_677_end_mask_0"), val = tensor([true, true, true])]; + tensor var_677 = slice_by_index(begin = var_677_begin_0, end = var_677_end_0, end_mask = var_677_end_mask_0, x = window_27)[name = tensor("op_677")]; + tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; + tensor window_29 = concat(axis = var_27, interleave = window_29_interleave_0, values = (var_677, var_674))[name = tensor("window_29")]; + tensor var_682_begin_0 = const()[name = tensor("op_682_begin_0"), val = tensor([0, 2, 0])]; + tensor var_682_end_0 = const()[name = tensor("op_682_end_0"), val = tensor([1, 3, 256])]; + tensor var_682_end_mask_0 = const()[name = tensor("op_682_end_mask_0"), val = tensor([true, false, true])]; + tensor var_682 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = x_15)[name = tensor("op_682")]; + tensor var_685_begin_0 = const()[name = tensor("op_685_begin_0"), val = tensor([0, 1, 0])]; + tensor var_685_end_0 = const()[name = tensor("op_685_end_0"), val = tensor([1, 16, 256])]; + tensor var_685_end_mask_0 = const()[name = tensor("op_685_end_mask_0"), val = tensor([true, true, true])]; + tensor var_685 = slice_by_index(begin = var_685_begin_0, end = var_685_end_0, end_mask = var_685_end_mask_0, x = window_29)[name = tensor("op_685")]; + tensor window_31_interleave_0 = const()[name = tensor("window_31_interleave_0"), val = tensor(false)]; + tensor window_31 = concat(axis = var_27, interleave = window_31_interleave_0, values = (var_685, var_682))[name = tensor("window_31")]; + tensor var_690_begin_0 = const()[name = tensor("op_690_begin_0"), val = tensor([0, 3, 0])]; + tensor var_690_end_0 = const()[name = tensor("op_690_end_0"), val = tensor([1, 4, 256])]; + tensor var_690_end_mask_0 = const()[name = tensor("op_690_end_mask_0"), val = tensor([true, false, true])]; + tensor var_690 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, x = x_15)[name = tensor("op_690")]; + tensor var_693_begin_0 = const()[name = tensor("op_693_begin_0"), val = tensor([0, 1, 0])]; + tensor var_693_end_0 = const()[name = tensor("op_693_end_0"), val = tensor([1, 16, 256])]; + tensor var_693_end_mask_0 = const()[name = tensor("op_693_end_mask_0"), val = tensor([true, true, true])]; + tensor var_693 = slice_by_index(begin = var_693_begin_0, end = var_693_end_0, end_mask = var_693_end_mask_0, x = window_31)[name = tensor("op_693")]; + tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; + tensor window_33 = concat(axis = var_27, interleave = window_33_interleave_0, values = (var_693, var_690))[name = tensor("window_33")]; + tensor var_698_begin_0 = const()[name = tensor("op_698_begin_0"), val = tensor([0, 4, 0])]; + tensor var_698_end_0 = const()[name = tensor("op_698_end_0"), val = tensor([1, 1, 256])]; + tensor var_698_end_mask_0 = const()[name = tensor("op_698_end_mask_0"), val = tensor([true, true, true])]; + tensor var_698 = slice_by_index(begin = var_698_begin_0, end = var_698_end_0, end_mask = var_698_end_mask_0, x = x_15)[name = tensor("op_698")]; + tensor var_701_begin_0 = const()[name = tensor("op_701_begin_0"), val = tensor([0, 1, 0])]; + tensor var_701_end_0 = const()[name = tensor("op_701_end_0"), val = tensor([1, 16, 256])]; + tensor var_701_end_mask_0 = const()[name = tensor("op_701_end_mask_0"), val = tensor([true, true, true])]; + tensor var_701 = slice_by_index(begin = var_701_begin_0, end = var_701_end_0, end_mask = var_701_end_mask_0, x = window_33)[name = tensor("op_701")]; + tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; + tensor window_35 = concat(axis = var_27, interleave = window_35_interleave_0, values = (var_701, var_698))[name = tensor("window_35")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_24, interleave = input_101_interleave_0, values = (window_27, window_29, window_31, window_33, window_35))[name = tensor("input_101")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_726_split_sizes_0 = const()[name = tensor("op_726_split_sizes_0"), val = tensor([256, 256])]; + tensor var_726_axis_0 = const()[name = tensor("op_726_axis_0"), val = tensor(1)]; + tensor var_726_0, tensor var_726_1 = split(axis = var_726_axis_0, split_sizes = var_726_split_sizes_0, x = inputs_23)[name = tensor("op_726")]; + tensor var_728 = sigmoid(x = var_726_1)[name = tensor("op_728")]; + tensor inputs_25 = mul(x = var_726_0, y = var_728)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([5, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([0, -1, 0])]; + tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([5, 16, 256])]; + tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_759 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, x = conv_out_5)[name = tensor("op_759")]; + tensor var_761_perm_0 = const()[name = tensor("op_761_perm_0"), val = tensor([1, 0, 2])]; + tensor var_761 = transpose(perm = var_761_perm_0, x = var_759)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_761)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_784 = const()[name = tensor("op_784"), val = tensor(0x1p-1)]; + tensor var_785 = mul(x = input_119, y = var_784)[name = tensor("op_785")]; + tensor input_121 = add(x = var_785, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_29, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_814 = const()[name = tensor("op_814"), val = tensor(0x1p-1)]; + tensor var_815 = mul(x = input_131, y = var_814)[name = tensor("op_815")]; + tensor input_133 = add(x = var_815, y = input_123)[name = tensor("input_133")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_29, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_829 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_830 = const()[name = tensor("op_830"), val = tensor([1, 5, 4, 64])]; + tensor var_831 = reshape(shape = var_830, x = var_829)[name = tensor("op_831")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_835 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_836 = const()[name = tensor("op_836"), val = tensor(0x1p-3)]; + tensor var_837 = mul(x = var_835, y = var_836)[name = tensor("op_837")]; + tensor var_838 = const()[name = tensor("op_838"), val = tensor([1, 5, 4, 64])]; + tensor var_839 = reshape(shape = var_838, x = var_837)[name = tensor("op_839")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_843 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_844 = const()[name = tensor("op_844"), val = tensor([1, 5, 4, 64])]; + tensor var_845 = reshape(shape = var_844, x = var_843)[name = tensor("op_845")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_839)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_831)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_855 = const()[name = tensor("op_855"), val = tensor([5, 1])]; + tensor var_856 = reshape(shape = var_855, x = sqrt_s_t_7)[name = tensor("op_856")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_856)[name = tensor("M_7")]; + tensor var_858 = mul(x = qk_7, y = M_7)[name = tensor("op_858")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_845)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_858, y = v_7)[name = tensor("inner_7")]; + tensor var_860_transpose_x_0 = const()[name = tensor("op_860_transpose_x_0"), val = tensor(false)]; + tensor var_860_transpose_y_0 = const()[name = tensor("op_860_transpose_y_0"), val = tensor(false)]; + tensor var_860 = matmul(transpose_x = var_860_transpose_x_0, transpose_y = var_860_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_860")]; + tensor var_861 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_861")]; + tensor var_862 = const()[name = tensor("op_862"), val = tensor([1, 1, 5, 1])]; + tensor var_863 = reshape(shape = var_862, x = var_861)[name = tensor("op_863")]; + tensor cross_7 = mul(x = var_860, y = var_863)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_866 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_866")]; + tensor var_868_transpose_x_1 = const()[name = tensor("op_868_transpose_x_1"), val = tensor(true)]; + tensor var_868_transpose_y_1 = const()[name = tensor("op_868_transpose_y_1"), val = tensor(false)]; + tensor var_868 = matmul(transpose_x = var_868_transpose_x_1, transpose_y = var_868_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_868")]; + tensor new_kv_unnorm_7 = add(x = var_866, y = var_868)[name = tensor("new_kv_unnorm_7")]; + tensor var_870 = const()[name = tensor("op_870"), val = tensor(0x1.4p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_870)[name = tensor("new_scale_7")]; + tensor var_872 = sqrt(x = new_scale_7)[name = tensor("op_872")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_872)[name = tensor("nkv_1")]; + tensor var_874_perm_0 = const()[name = tensor("op_874_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_874 = transpose(perm = var_874_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_18, x = var_874)[name = tensor("out_21")]; + tensor var_878 = const()[name = tensor("op_878"), val = tensor([1, 5, 256])]; + tensor out_23 = reshape(shape = var_878, x = out_21)[name = tensor("out_23")]; + tensor var_880 = silu(x = input_137)[name = tensor("op_880")]; + tensor input_139 = mul(x = var_880, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_37_begin_0 = const()[name = tensor("window_37_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_37_end_0 = const()[name = tensor("window_37_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_37_end_mask_0 = const()[name = tensor("window_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_37_squeeze_mask_0 = const()[name = tensor("window_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_37 = slice_by_index(begin = window_37_begin_0, end = window_37_end_0, end_mask = window_37_end_mask_0, squeeze_mask = window_37_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_37")]; + tensor var_888_begin_0 = const()[name = tensor("op_888_begin_0"), val = tensor([0, 0, 0])]; + tensor var_888_end_0 = const()[name = tensor("op_888_end_0"), val = tensor([1, 1, 256])]; + tensor var_888_end_mask_0 = const()[name = tensor("op_888_end_mask_0"), val = tensor([true, false, true])]; + tensor var_888 = slice_by_index(begin = var_888_begin_0, end = var_888_end_0, end_mask = var_888_end_mask_0, x = x_21)[name = tensor("op_888")]; + tensor var_891_begin_0 = const()[name = tensor("op_891_begin_0"), val = tensor([0, 1, 0])]; + tensor var_891_end_0 = const()[name = tensor("op_891_end_0"), val = tensor([1, 16, 256])]; + tensor var_891_end_mask_0 = const()[name = tensor("op_891_end_mask_0"), val = tensor([true, true, true])]; + tensor var_891 = slice_by_index(begin = var_891_begin_0, end = var_891_end_0, end_mask = var_891_end_mask_0, x = window_37)[name = tensor("op_891")]; + tensor window_39_interleave_0 = const()[name = tensor("window_39_interleave_0"), val = tensor(false)]; + tensor window_39 = concat(axis = var_27, interleave = window_39_interleave_0, values = (var_891, var_888))[name = tensor("window_39")]; + tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 1, 0])]; + tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 2, 256])]; + tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, false, true])]; + tensor var_896 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = x_21)[name = tensor("op_896")]; + tensor var_899_begin_0 = const()[name = tensor("op_899_begin_0"), val = tensor([0, 1, 0])]; + tensor var_899_end_0 = const()[name = tensor("op_899_end_0"), val = tensor([1, 16, 256])]; + tensor var_899_end_mask_0 = const()[name = tensor("op_899_end_mask_0"), val = tensor([true, true, true])]; + tensor var_899 = slice_by_index(begin = var_899_begin_0, end = var_899_end_0, end_mask = var_899_end_mask_0, x = window_39)[name = tensor("op_899")]; + tensor window_41_interleave_0 = const()[name = tensor("window_41_interleave_0"), val = tensor(false)]; + tensor window_41 = concat(axis = var_27, interleave = window_41_interleave_0, values = (var_899, var_896))[name = tensor("window_41")]; + tensor var_904_begin_0 = const()[name = tensor("op_904_begin_0"), val = tensor([0, 2, 0])]; + tensor var_904_end_0 = const()[name = tensor("op_904_end_0"), val = tensor([1, 3, 256])]; + tensor var_904_end_mask_0 = const()[name = tensor("op_904_end_mask_0"), val = tensor([true, false, true])]; + tensor var_904 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = x_21)[name = tensor("op_904")]; + tensor var_907_begin_0 = const()[name = tensor("op_907_begin_0"), val = tensor([0, 1, 0])]; + tensor var_907_end_0 = const()[name = tensor("op_907_end_0"), val = tensor([1, 16, 256])]; + tensor var_907_end_mask_0 = const()[name = tensor("op_907_end_mask_0"), val = tensor([true, true, true])]; + tensor var_907 = slice_by_index(begin = var_907_begin_0, end = var_907_end_0, end_mask = var_907_end_mask_0, x = window_41)[name = tensor("op_907")]; + tensor window_43_interleave_0 = const()[name = tensor("window_43_interleave_0"), val = tensor(false)]; + tensor window_43 = concat(axis = var_27, interleave = window_43_interleave_0, values = (var_907, var_904))[name = tensor("window_43")]; + tensor var_912_begin_0 = const()[name = tensor("op_912_begin_0"), val = tensor([0, 3, 0])]; + tensor var_912_end_0 = const()[name = tensor("op_912_end_0"), val = tensor([1, 4, 256])]; + tensor var_912_end_mask_0 = const()[name = tensor("op_912_end_mask_0"), val = tensor([true, false, true])]; + tensor var_912 = slice_by_index(begin = var_912_begin_0, end = var_912_end_0, end_mask = var_912_end_mask_0, x = x_21)[name = tensor("op_912")]; + tensor var_915_begin_0 = const()[name = tensor("op_915_begin_0"), val = tensor([0, 1, 0])]; + tensor var_915_end_0 = const()[name = tensor("op_915_end_0"), val = tensor([1, 16, 256])]; + tensor var_915_end_mask_0 = const()[name = tensor("op_915_end_mask_0"), val = tensor([true, true, true])]; + tensor var_915 = slice_by_index(begin = var_915_begin_0, end = var_915_end_0, end_mask = var_915_end_mask_0, x = window_43)[name = tensor("op_915")]; + tensor window_45_interleave_0 = const()[name = tensor("window_45_interleave_0"), val = tensor(false)]; + tensor window_45 = concat(axis = var_27, interleave = window_45_interleave_0, values = (var_915, var_912))[name = tensor("window_45")]; + tensor var_920_begin_0 = const()[name = tensor("op_920_begin_0"), val = tensor([0, 4, 0])]; + tensor var_920_end_0 = const()[name = tensor("op_920_end_0"), val = tensor([1, 1, 256])]; + tensor var_920_end_mask_0 = const()[name = tensor("op_920_end_mask_0"), val = tensor([true, true, true])]; + tensor var_920 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = x_21)[name = tensor("op_920")]; + tensor var_923_begin_0 = const()[name = tensor("op_923_begin_0"), val = tensor([0, 1, 0])]; + tensor var_923_end_0 = const()[name = tensor("op_923_end_0"), val = tensor([1, 16, 256])]; + tensor var_923_end_mask_0 = const()[name = tensor("op_923_end_mask_0"), val = tensor([true, true, true])]; + tensor var_923 = slice_by_index(begin = var_923_begin_0, end = var_923_end_0, end_mask = var_923_end_mask_0, x = window_45)[name = tensor("op_923")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_27, interleave = window_interleave_0, values = (var_923, var_920))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_24, interleave = input_141_interleave_0, values = (window_39, window_41, window_43, window_45, window))[name = tensor("input_141")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_948_split_sizes_0 = const()[name = tensor("op_948_split_sizes_0"), val = tensor([256, 256])]; + tensor var_948_axis_0 = const()[name = tensor("op_948_axis_0"), val = tensor(1)]; + tensor var_948_0, tensor var_948_1 = split(axis = var_948_axis_0, split_sizes = var_948_split_sizes_0, x = inputs_33)[name = tensor("op_948")]; + tensor var_950 = sigmoid(x = var_948_1)[name = tensor("op_950")]; + tensor inputs_35 = mul(x = var_948_0, y = var_950)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([5, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_981_begin_0 = const()[name = tensor("op_981_begin_0"), val = tensor([0, -1, 0])]; + tensor var_981_end_0 = const()[name = tensor("op_981_end_0"), val = tensor([5, 16, 256])]; + tensor var_981_end_mask_0 = const()[name = tensor("op_981_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_981 = slice_by_index(begin = var_981_begin_0, end = var_981_end_0, end_mask = var_981_end_mask_0, x = conv_out_7)[name = tensor("op_981")]; + tensor var_983_perm_0 = const()[name = tensor("op_983_perm_0"), val = tensor([1, 0, 2])]; + tensor var_983 = transpose(perm = var_983_perm_0, x = var_981)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_983)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_1006 = const()[name = tensor("op_1006"), val = tensor(0x1p-1)]; + tensor var_1007 = mul(x = input_159, y = var_1006)[name = tensor("op_1007")]; + tensor input_161 = add(x = var_1007, y = input_151)[name = tensor("input_161")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_29, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_21, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1025_begin_0 = const()[name = tensor("op_1025_begin_0"), val = tensor([0, 0, 5])]; + tensor var_1025_end_0 = const()[name = tensor("op_1025_end_0"), val = tensor([1, 256, 23])]; + tensor var_1025_end_mask_0 = const()[name = tensor("op_1025_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1025_begin_0, end = var_1025_end_0, end_mask = var_1025_end_mask_0, x = cat)[name = tensor("op_1025")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1028 = reduce_l2_norm(axes = var_1027, keep_dims = var_30, x = input_163)[name = tensor("op_1028")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_1028)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_1032_axis_0 = const()[name = tensor("op_1032_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1032_axis_0, values = (var_207, var_429, var_651, nkv_1))[name = tensor("op_1032")]; + tensor var_1034_axis_0 = const()[name = tensor("op_1034_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1034_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1034")]; + tensor var_1036_axis_0 = const()[name = tensor("op_1036_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1036_axis_0, values = (window_11, window_23, window_35, window))[name = tensor("op_1036")]; + tensor var_1045 = const()[name = tensor("op_1045"), val = tensor(0x1.5798eep-27)]; + tensor var_1050 = const()[name = tensor("op_1050"), val = tensor(0x1.4f8b58p-17)]; + tensor var_1052 = const()[name = tensor("op_1052"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_1053 = const()[name = tensor("op_1053"), val = tensor(true)]; + tensor var_1055 = const()[name = tensor("op_1055"), val = tensor(0x1p+0)]; + tensor var_1059 = const()[name = tensor("op_1059"), val = tensor(-1)]; + tensor var_1065 = const()[name = tensor("op_1065"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395712)))]; + tensor var_1127_axes_0 = const()[name = tensor("op_1127_axes_0"), val = tensor([2])]; + tensor var_1127 = expand_dims(axes = var_1127_axes_0, x = emb)[name = tensor("op_1127")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1127)[name = tensor("emb_exp")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_1059, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; + tensor var_1135_perm_0 = const()[name = tensor("op_1135_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1139 = const()[name = tensor("op_1139"), val = tensor([12, 5, 256])]; + tensor var_1135 = transpose(perm = var_1135_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1139, x = var_1135)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1147 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([12, 5, 4, 64])]; + tensor var_1149 = reshape(shape = var_1148, x = var_1147)[name = tensor("op_1149")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1153 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor(0x1p-3)]; + tensor var_1155 = mul(x = var_1153, y = var_1154)[name = tensor("op_1155")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([12, 5, 4, 64])]; + tensor var_1157 = reshape(shape = var_1156, x = var_1155)[name = tensor("op_1157")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1161 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor([12, 5, 4, 64])]; + tensor var_1163 = reshape(shape = var_1162, x = var_1161)[name = tensor("op_1163")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_1065, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_1055, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1157)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1149)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 5])]; + tensor var_1176 = reshape(shape = var_1175, x = valid_mask)[name = tensor("op_1176")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1176)[name = tensor("causal_with_valid_1")]; + tensor var_1178 = const()[name = tensor("op_1178"), val = tensor([5, 1])]; + tensor var_1179 = reshape(shape = var_1178, x = sqrt_s_t_9)[name = tensor("op_1179")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1179)[name = tensor("M_9")]; + tensor var_1181 = mul(x = qk_9, y = M_9)[name = tensor("op_1181")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1163)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1181, y = v_9)[name = tensor("inner_9")]; + tensor var_1183_transpose_x_0 = const()[name = tensor("op_1183_transpose_x_0"), val = tensor(false)]; + tensor var_1183_transpose_y_0 = const()[name = tensor("op_1183_transpose_y_0"), val = tensor(false)]; + tensor var_1183 = matmul(transpose_x = var_1183_transpose_x_0, transpose_y = var_1183_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1183")]; + tensor var_1184 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1184")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 5, 1])]; + tensor var_1186 = reshape(shape = var_1185, x = var_1184)[name = tensor("op_1186")]; + tensor cross_9 = mul(x = var_1183, y = var_1186)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1, 1, 5, 1])]; + tensor var_1190 = reshape(shape = var_1189, x = valid_mask)[name = tensor("op_1190")]; + tensor v_masked_1 = mul(x = v_9, y = var_1190)[name = tensor("v_masked_1")]; + tensor var_1192 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1192")]; + tensor var_1194_transpose_x_1 = const()[name = tensor("op_1194_transpose_x_1"), val = tensor(true)]; + tensor var_1194_transpose_y_1 = const()[name = tensor("op_1194_transpose_y_1"), val = tensor(false)]; + tensor var_1194 = matmul(transpose_x = var_1194_transpose_x_1, transpose_y = var_1194_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1194")]; + tensor new_kv_unnorm_9 = add(x = var_1192, y = var_1194)[name = tensor("new_kv_unnorm_9")]; + tensor var_1196_keep_dims_0 = const()[name = tensor("op_1196_keep_dims_0"), val = tensor(false)]; + tensor var_1196 = reduce_sum(keep_dims = var_1196_keep_dims_0, x = valid_mask)[name = tensor("op_1196")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1])]; + tensor var_1198 = reshape(shape = var_1197, x = var_1196)[name = tensor("op_1198")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1198)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_1055, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1202 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1202")]; + tensor var_1203_perm_0 = const()[name = tensor("op_1203_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1203 = transpose(perm = var_1203_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1052, x = var_1203)[name = tensor("out_27")]; + tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([12, 5, 256])]; + tensor out_29 = reshape(shape = var_1207, x = out_27)[name = tensor("out_29")]; + tensor var_1209 = silu(x = input_169)[name = tensor("op_1209")]; + tensor input_171 = mul(x = var_1209, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1050, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1219 = const()[name = tensor("op_1219"), val = tensor([1, 12, 5, 256])]; + tensor var_1220 = reshape(shape = var_1219, x = xt_1)[name = tensor("op_1220")]; + tensor var_1221_perm_0 = const()[name = tensor("op_1221_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([5, 12, 256])]; + tensor var_1221 = transpose(perm = var_1221_perm_0, x = var_1220)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1224, x = var_1221)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1247 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 5, 3, 256])]; + tensor var_1249 = reshape(shape = concat_1, x = var_1247)[name = tensor("op_1249")]; + tensor var_1250_axes_0 = const()[name = tensor("op_1250_axes_0"), val = tensor([0])]; + tensor var_1250 = expand_dims(axes = var_1250_axes_0, x = var_1249)[name = tensor("op_1250")]; + tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1252_axes_0 = const()[name = tensor("op_1252_axes_0"), val = tensor([-2])]; + tensor var_1251 = transpose(perm = var_1251_perm_0, x = var_1250)[name = tensor("transpose_21")]; + tensor var_1252 = squeeze(axes = var_1252_axes_0, x = var_1251)[name = tensor("op_1252")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 5, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1252)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 5, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1252)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 5, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1252)[name = tensor("v_11")]; + tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([12, 20, 64])]; + tensor var_1261 = reshape(shape = var_1260, x = q_11)[name = tensor("op_1261")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1267 = const()[name = tensor("op_1267"), val = tensor([12, 20, 64])]; + tensor var_1268 = reshape(shape = var_1267, x = k_11)[name = tensor("op_1268")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([12, 20, 64])]; + tensor var_1275 = reshape(shape = var_1274, x = v_11)[name = tensor("op_1275")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([5, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1261)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1278, x = q_13)[name = tensor("q_15")]; + tensor var_1280 = const()[name = tensor("op_1280"), val = tensor([5, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1268)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1280, x = k_13)[name = tensor("k_15")]; + tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([5, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1275)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1282, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1285 = const()[name = tensor("op_1285"), val = tensor([2, 0, 1, 3])]; + tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([60, 256])]; + tensor var_1286 = transpose(perm = var_1285, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1290, x = var_1286)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([12, 5, 256])]; + tensor attn_output_7 = reshape(shape = var_1294, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1050, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1050, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 5, 12, 256])]; + tensor x_31 = reshape(shape = var_1314, x = xt_3)[name = tensor("x_31")]; + tensor var_1316_perm_0 = const()[name = tensor("op_1316_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1320 = const()[name = tensor("op_1320"), val = tensor([12, 5, 256])]; + tensor var_1316 = transpose(perm = var_1316_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1320, x = var_1316)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1328 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([12, 5, 4, 64])]; + tensor var_1330 = reshape(shape = var_1329, x = var_1328)[name = tensor("op_1330")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1334 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor(0x1p-3)]; + tensor var_1336 = mul(x = var_1334, y = var_1335)[name = tensor("op_1336")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([12, 5, 4, 64])]; + tensor var_1338 = reshape(shape = var_1337, x = var_1336)[name = tensor("op_1338")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1342 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([12, 5, 4, 64])]; + tensor var_1344 = reshape(shape = var_1343, x = var_1342)[name = tensor("op_1344")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_1055, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1338)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1330)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1359 = const()[name = tensor("op_1359"), val = tensor([5, 1])]; + tensor var_1360 = reshape(shape = var_1359, x = sqrt_s_t)[name = tensor("op_1360")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1360)[name = tensor("M")]; + tensor var_1362 = mul(x = qk, y = M)[name = tensor("op_1362")]; + tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1344)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1362, y = v_17)[name = tensor("inner")]; + tensor var_1364_transpose_x_0 = const()[name = tensor("op_1364_transpose_x_0"), val = tensor(false)]; + tensor var_1364_transpose_y_0 = const()[name = tensor("op_1364_transpose_y_0"), val = tensor(false)]; + tensor var_1364 = matmul(transpose_x = var_1364_transpose_x_0, transpose_y = var_1364_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1364")]; + tensor var_1365 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1365")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 5, 1])]; + tensor var_1367 = reshape(shape = var_1366, x = var_1365)[name = tensor("op_1367")]; + tensor cross = mul(x = var_1364, y = var_1367)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1190)[name = tensor("v_masked")]; + tensor var_1373 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1373")]; + tensor var_1375_transpose_x_1 = const()[name = tensor("op_1375_transpose_x_1"), val = tensor(true)]; + tensor var_1375_transpose_y_1 = const()[name = tensor("op_1375_transpose_y_1"), val = tensor(false)]; + tensor var_1375 = matmul(transpose_x = var_1375_transpose_x_1, transpose_y = var_1375_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1375")]; + tensor new_kv_unnorm = add(x = var_1373, y = var_1375)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1198)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_1055, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1384_perm_0 = const()[name = tensor("op_1384_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1384 = transpose(perm = var_1384_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1052, x = var_1384)[name = tensor("out_33")]; + tensor var_1388 = const()[name = tensor("op_1388"), val = tensor([12, 5, 256])]; + tensor out = reshape(shape = var_1388, x = out_33)[name = tensor("out")]; + tensor var_1390 = silu(x = input_187)[name = tensor("op_1390")]; + tensor input_189 = mul(x = var_1390, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1050, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 12, 5, 256])]; + tensor var_1401 = reshape(shape = var_1400, x = xt_5)[name = tensor("op_1401")]; + tensor var_1402_perm_0 = const()[name = tensor("op_1402_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1405 = const()[name = tensor("op_1405"), val = tensor([5, 12, 256])]; + tensor var_1402 = transpose(perm = var_1402_perm_0, x = var_1401)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1405, x = var_1402)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1428 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 5, 3, 256])]; + tensor var_1430 = reshape(shape = concat_2, x = var_1428)[name = tensor("op_1430")]; + tensor var_1431_axes_0 = const()[name = tensor("op_1431_axes_0"), val = tensor([0])]; + tensor var_1431 = expand_dims(axes = var_1431_axes_0, x = var_1430)[name = tensor("op_1431")]; + tensor var_1432_perm_0 = const()[name = tensor("op_1432_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1433_axes_0 = const()[name = tensor("op_1433_axes_0"), val = tensor([-2])]; + tensor var_1432 = transpose(perm = var_1432_perm_0, x = var_1431)[name = tensor("transpose_8")]; + tensor var_1433 = squeeze(axes = var_1433_axes_0, x = var_1432)[name = tensor("op_1433")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 5, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1433)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 5, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1433)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 5, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1433)[name = tensor("v_19")]; + tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([12, 20, 64])]; + tensor var_1442 = reshape(shape = var_1441, x = q_19)[name = tensor("op_1442")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([12, 20, 64])]; + tensor var_1449 = reshape(shape = var_1448, x = k_19)[name = tensor("op_1449")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([12, 20, 64])]; + tensor var_1456 = reshape(shape = var_1455, x = v_19)[name = tensor("op_1456")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([5, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1442)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1459, x = q_21)[name = tensor("q")]; + tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([5, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1449)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1461, x = k_21)[name = tensor("k")]; + tensor var_1463 = const()[name = tensor("op_1463"), val = tensor([5, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1456)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1463, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1466 = const()[name = tensor("op_1466"), val = tensor([2, 0, 1, 3])]; + tensor var_1471 = const()[name = tensor("op_1471"), val = tensor([60, 256])]; + tensor var_1467 = transpose(perm = var_1466, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1471, x = var_1467)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1475 = const()[name = tensor("op_1475"), val = tensor([12, 5, 256])]; + tensor attn_output = reshape(shape = var_1475, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1050, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1050, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1495 = const()[name = tensor("op_1495"), val = tensor([1, 5, 12, 256])]; + tensor input = reshape(shape = var_1495, x = xt)[name = tensor("input")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([-1])]; + tensor var_1498 = reduce_l2_norm(axes = var_1497, keep_dims = var_1053, x = input)[name = tensor("op_1498")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_1045, beta = const_42, x = var_1498)[name = tensor("clip_5")]; + tensor var_1500 = real_div(x = input, y = clip_5)[name = tensor("op_1500")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([5, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([5, 256, 12])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1500)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 5, 12])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 5, 11])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1504")]; + tensor var_1506_axis_0 = const()[name = tensor("op_1506_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1506_axis_0, values = (var_1202, nkv))[name = tensor("op_1506")]; + tensor var_1508_axis_0 = const()[name = tensor("op_1508_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1508_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1508")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file diff --git a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/weights/weight.bin b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..5be300709dc9ecfc319c29b8e5e0fd5409fdf07f --- /dev/null +++ b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76ce7934e2cd91a0effb70dfa90fb24edb66bc9a764b290ee2065b896721b3cb +size 44457216 diff --git a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..55956bf9724e5d26a5683d705abec5e007070969 --- /dev/null +++ b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:caa9e47ded51d738809bca4aef08aede86aea03f444055505cb42b1a9539bb14 +size 196629 diff --git a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..5be300709dc9ecfc319c29b8e5e0fd5409fdf07f --- /dev/null +++ b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76ce7934e2cd91a0effb70dfa90fb24edb66bc9a764b290ee2065b896721b3cb +size 44457216 diff --git a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Manifest.json b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..eb273aebdeb752ae1ebf3dffb84bfed3c48e15a9 --- /dev/null +++ b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "36D483E8-31EB-4224-A050-837C09E85C85": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "F67F7E02-4DC3-4166-A52A-2D9FE21F41D8": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "36D483E8-31EB-4224-A050-837C09E85C85" +}