program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}})]
{
    func tower_10s<ios18>(tensor<fp32, [1, 1, 130, 130]> attn_mask, tensor<fp32, [1, 128, 1000]> audios) {
            tensor<int32, [3]> var_25_begin_0 = const()[name = string("op_25_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_25_end_0 = const()[name = string("op_25_end_0"), val = tensor<int32, [3]>([1, 128, 1000])];
            tensor<bool, [3]> var_25_end_mask_0 = const()[name = string("op_25_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> var_25_squeeze_mask_0 = const()[name = string("op_25_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            string audios_to_fp16_dtype_0 = const()[name = string("audios_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1000]> audios_to_fp16 = cast(dtype = audios_to_fp16_dtype_0, x = audios)[name = string("cast_1")];
            tensor<fp16, [128, 1000]> var_25_cast_fp16 = slice_by_index(begin = var_25_begin_0, end = var_25_end_0, end_mask = var_25_end_mask_0, squeeze_mask = var_25_squeeze_mask_0, x = audios_to_fp16)[name = string("op_25_cast_fp16")];
            tensor<int32, [2]> x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor<int32, [2]>([1, 0])];
            tensor<int32, [3]> var_27 = const()[name = string("op_27"), val = tensor<int32, [3]>([10, 100, 128])];
            tensor<fp16, [1000, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = var_25_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [10, 100, 128]> x_3_cast_fp16 = reshape(shape = var_27, x = x_1_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<int32, [3]> var_29 = const()[name = string("op_29"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_1_axes_0 = const()[name = string("input_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [10, 128, 100]> var_30_cast_fp16 = transpose(perm = var_29, x = x_3_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [10, 1, 128, 100]> input_1_cast_fp16 = expand_dims(axes = input_1_axes_0, x = var_30_cast_fp16)[name = string("input_1_cast_fp16")];
            string var_38_pad_type_0 = const()[name = string("op_38_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_38_pad_0 = const()[name = string("op_38_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> var_38_strides_0 = const()[name = string("op_38_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_38_dilations_0 = const()[name = string("op_38_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_38_groups_0 = const()[name = string("op_38_groups_0"), val = int32(1)];
            tensor<fp16, [480, 1, 3, 3]> frontend_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("frontend_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [480, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5056))), zero_point = tensor<int8, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4480)))];
            tensor<fp16, [480]> frontend_conv1_bias_to_fp16 = const()[name = string("frontend_conv1_bias_to_fp16"), val = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6080)))];
            tensor<fp16, [10, 480, 64, 50]> var_38_cast_fp16 = conv(bias = frontend_conv1_bias_to_fp16, dilations = var_38_dilations_0, groups = var_38_groups_0, pad = var_38_pad_0, pad_type = var_38_pad_type_0, strides = var_38_strides_0, weight = frontend_conv1_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("op_38_cast_fp16")];
            string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("EXACT")];
            tensor<fp16, [10, 480, 64, 50]> input_3_cast_fp16 = gelu(mode = input_3_mode_0, x = var_38_cast_fp16)[name = string("input_3_cast_fp16")];
            string var_46_pad_type_0 = const()[name = string("op_46_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_46_pad_0 = const()[name = string("op_46_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> var_46_strides_0 = const()[name = string("op_46_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_46_dilations_0 = const()[name = string("op_46_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_46_groups_0 = const()[name = string("op_46_groups_0"), val = int32(1)];
            tensor<fp16, [480, 480, 3, 3]> frontend_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("frontend_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [480, 480, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7104))), scale = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2080768))), zero_point = tensor<int8, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4480)))];
            tensor<fp16, [480]> frontend_conv2_bias_to_fp16 = const()[name = string("frontend_conv2_bias_to_fp16"), val = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2081792)))];
            tensor<fp16, [10, 480, 32, 25]> var_46_cast_fp16 = conv(bias = frontend_conv2_bias_to_fp16, dilations = var_46_dilations_0, groups = var_46_groups_0, pad = var_46_pad_0, pad_type = var_46_pad_type_0, strides = var_46_strides_0, weight = frontend_conv2_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("op_46_cast_fp16")];
            string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("EXACT")];
            tensor<fp16, [10, 480, 32, 25]> input_5_cast_fp16 = gelu(mode = input_5_mode_0, x = var_46_cast_fp16)[name = string("input_5_cast_fp16")];
            string var_54_pad_type_0 = const()[name = string("op_54_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_54_pad_0 = const()[name = string("op_54_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> var_54_strides_0 = const()[name = string("op_54_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_54_dilations_0 = const()[name = string("op_54_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_54_groups_0 = const()[name = string("op_54_groups_0"), val = int32(1)];
            tensor<fp16, [480, 480, 3, 3]> frontend_conv3_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("frontend_conv3_weight_to_fp16_quantized"), quantized_data = tensor<int8, [480, 480, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2082816))), scale = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4156480))), zero_point = tensor<int8, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4480)))];
            tensor<fp16, [480]> frontend_conv3_bias_to_fp16 = const()[name = string("frontend_conv3_bias_to_fp16"), val = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4157504)))];
            tensor<fp16, [10, 480, 16, 13]> var_54_cast_fp16 = conv(bias = frontend_conv3_bias_to_fp16, dilations = var_54_dilations_0, groups = var_54_groups_0, pad = var_54_pad_0, pad_type = var_54_pad_type_0, strides = var_54_strides_0, weight = frontend_conv3_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("op_54_cast_fp16")];
            string x_5_mode_0 = const()[name = string("x_5_mode_0"), val = string("EXACT")];
            tensor<fp16, [10, 480, 16, 13]> x_5_cast_fp16 = gelu(mode = x_5_mode_0, x = var_54_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<int32, [4]> var_56 = const()[name = string("op_56"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [3]> var_58 = const()[name = string("op_58"), val = tensor<int32, [3]>([10, 13, 7680])];
            tensor<fp16, [10, 13, 480, 16]> var_57_cast_fp16 = transpose(perm = var_56, x = x_5_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [10, 13, 7680]> input_7_cast_fp16 = reshape(shape = var_58, x = var_57_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [896, 7680]> frontend_conv_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("frontend_conv_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 7680]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4158528))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11040832))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11042688)))];
            tensor<fp16, [10, 13, 896]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = frontend_conv_out_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<fp16, [1, 13, 896]> frontend_pos_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("frontend_pos_to_fp16_quantized"), quantized_data = tensor<int8, [1, 13, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11044544))), scale = fp16(0x1.02p-7), zero_point = int8(0)];
            tensor<fp16, [10, 13, 896]> x_cast_fp16 = add(x = linear_0_cast_fp16, y = frontend_pos_to_fp16_quantized)[name = string("x_cast_fp16")];
            tensor<int32, [2]> var_63 = const()[name = string("op_63"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 896]> input_9_cast_fp16 = reshape(shape = var_63, x = x_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<int32, [1]> input_11_axes_0 = const()[name = string("input_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_0_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_0_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11056256)))];
            tensor<fp16, [896]> trunk_layers_0_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_0_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11058112)))];
            fp16 var_67_to_fp16 = const()[name = string("op_67_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [130, 896]> input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = trunk_layers_0_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_0_attn_ln_weight_to_fp16, x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_0_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11059968))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11862848))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_0_q_bias_to_fp16 = const()[name = string("trunk_layers_0_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11864704)))];
            tensor<fp16, [130, 896]> linear_1_cast_fp16 = linear(bias = trunk_layers_0_q_bias_to_fp16, weight = trunk_layers_0_q_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [3]> var_141 = const()[name = string("op_141"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_142_cast_fp16 = reshape(shape = var_141, x = linear_1_cast_fp16)[name = string("op_142_cast_fp16")];
            tensor<int32, [3]> var_143_perm_0 = const()[name = string("op_143_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_1_axes_0 = const()[name = string("q_1_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_143_cast_fp16 = transpose(perm = var_143_perm_0, x = var_142_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 14, 130, 64]> q_1_cast_fp16 = expand_dims(axes = q_1_axes_0, x = var_143_cast_fp16)[name = string("q_1_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_0_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11866560))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12669440))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_0_k_bias_to_fp16 = const()[name = string("trunk_layers_0_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12671296)))];
            tensor<fp16, [130, 896]> linear_2_cast_fp16 = linear(bias = trunk_layers_0_k_bias_to_fp16, weight = trunk_layers_0_k_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [3]> var_148 = const()[name = string("op_148"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_149_cast_fp16 = reshape(shape = var_148, x = linear_2_cast_fp16)[name = string("op_149_cast_fp16")];
            tensor<int32, [3]> var_150_perm_0 = const()[name = string("op_150_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_1_axes_0 = const()[name = string("k_1_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_150_cast_fp16 = transpose(perm = var_150_perm_0, x = var_149_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 14, 130, 64]> k_1_cast_fp16 = expand_dims(axes = k_1_axes_0, x = var_150_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_0_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12673152))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13476032))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_0_v_bias_to_fp16 = const()[name = string("trunk_layers_0_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13477888)))];
            tensor<fp16, [130, 896]> linear_3_cast_fp16 = linear(bias = trunk_layers_0_v_bias_to_fp16, weight = trunk_layers_0_v_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<int32, [3]> var_155 = const()[name = string("op_155"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_156_cast_fp16 = reshape(shape = var_155, x = linear_3_cast_fp16)[name = string("op_156_cast_fp16")];
            tensor<int32, [3]> var_157_perm_0 = const()[name = string("op_157_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_1_axes_0 = const()[name = string("v_1_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_157_cast_fp16 = transpose(perm = var_157_perm_0, x = var_156_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 14, 130, 64]> v_1_cast_fp16 = expand_dims(axes = v_1_axes_0, x = var_157_cast_fp16)[name = string("v_1_cast_fp16")];
            fp16 mul_0_y_0_to_fp16 = const()[name = string("mul_0_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_0_cast_fp16 = mul(x = q_1_cast_fp16, y = mul_0_y_0_to_fp16)[name = string("mul_0_cast_fp16")];
            bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(true)];
            bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_0_cast_fp16, y = k_1_cast_fp16)[name = string("matmul_0_cast_fp16")];
            string attn_mask_to_fp16_dtype_0 = const()[name = string("attn_mask_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1, 130, 130]> attn_mask_to_fp16 = cast(dtype = attn_mask_to_fp16_dtype_0, x = attn_mask)[name = string("cast_0")];
            tensor<fp16, [1, 14, 130, 130]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = attn_mask_to_fp16)[name = string("add_0_cast_fp16")];
            int32 softmax_0_axis_0 = const()[name = string("softmax_0_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = string("softmax_0_cast_fp16")];
            bool a_1_transpose_x_0 = const()[name = string("a_1_transpose_x_0"), val = bool(false)];
            bool a_1_transpose_y_0 = const()[name = string("a_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_1_cast_fp16 = matmul(transpose_x = a_1_transpose_x_0, transpose_y = a_1_transpose_y_0, x = softmax_0_cast_fp16, y = v_1_cast_fp16)[name = string("a_1_cast_fp16")];
            tensor<int32, [1]> var_160_axes_0 = const()[name = string("op_160_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_160_cast_fp16 = squeeze(axes = var_160_axes_0, x = a_1_cast_fp16)[name = string("op_160_cast_fp16")];
            tensor<int32, [3]> var_161_perm_0 = const()[name = string("op_161_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_162 = const()[name = string("op_162"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_161_cast_fp16 = transpose(perm = var_161_perm_0, x = var_160_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [130, 896]> input_13_cast_fp16 = reshape(shape = var_162, x = var_161_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_0_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13479744))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14282624))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_0_out_bias_to_fp16 = const()[name = string("trunk_layers_0_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14284480)))];
            tensor<fp16, [130, 896]> linear_4_cast_fp16 = linear(bias = trunk_layers_0_out_bias_to_fp16, weight = trunk_layers_0_out_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [130, 896]> input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<int32, [1]> input_17_axes_0 = const()[name = string("input_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_0_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_0_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14286336)))];
            tensor<fp16, [896]> trunk_layers_0_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_0_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14288192)))];
            tensor<fp16, [130, 896]> input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = trunk_layers_0_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_0_mlp_ln_weight_to_fp16, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_0_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14290048))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17505024))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_0_fc1_bias_to_fp16 = const()[name = string("trunk_layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17512256)))];
            tensor<fp16, [130, 3584]> linear_5_cast_fp16 = linear(bias = trunk_layers_0_fc1_bias_to_fp16, weight = trunk_layers_0_fc1_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_5_cast_fp16")];
            string input_19_mode_0 = const()[name = string("input_19_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = linear_5_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_0_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17519488))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20730816))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_0_fc2_bias_to_fp16 = const()[name = string("trunk_layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20732672)))];
            tensor<fp16, [130, 896]> linear_6_cast_fp16 = linear(bias = trunk_layers_0_fc2_bias_to_fp16, weight = trunk_layers_0_fc2_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [130, 896]> input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<int32, [1]> input_23_axes_0 = const()[name = string("input_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_1_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_1_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20734528)))];
            tensor<fp16, [896]> trunk_layers_1_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_1_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20736384)))];
            tensor<fp16, [130, 896]> input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = trunk_layers_1_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_1_attn_ln_weight_to_fp16, x = input_21_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_1_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20738240))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21541120))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_1_q_bias_to_fp16 = const()[name = string("trunk_layers_1_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21542976)))];
            tensor<fp16, [130, 896]> linear_7_cast_fp16 = linear(bias = trunk_layers_1_q_bias_to_fp16, weight = trunk_layers_1_q_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [3]> var_196 = const()[name = string("op_196"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_197_cast_fp16 = reshape(shape = var_196, x = linear_7_cast_fp16)[name = string("op_197_cast_fp16")];
            tensor<int32, [3]> var_198_perm_0 = const()[name = string("op_198_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_3_axes_0 = const()[name = string("q_3_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_198_cast_fp16 = transpose(perm = var_198_perm_0, x = var_197_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 14, 130, 64]> q_3_cast_fp16 = expand_dims(axes = q_3_axes_0, x = var_198_cast_fp16)[name = string("q_3_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_1_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21544832))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22347712))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_1_k_bias_to_fp16 = const()[name = string("trunk_layers_1_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22349568)))];
            tensor<fp16, [130, 896]> linear_8_cast_fp16 = linear(bias = trunk_layers_1_k_bias_to_fp16, weight = trunk_layers_1_k_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [3]> var_203 = const()[name = string("op_203"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_204_cast_fp16 = reshape(shape = var_203, x = linear_8_cast_fp16)[name = string("op_204_cast_fp16")];
            tensor<int32, [3]> var_205_perm_0 = const()[name = string("op_205_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_3_axes_0 = const()[name = string("k_3_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_205_cast_fp16 = transpose(perm = var_205_perm_0, x = var_204_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 14, 130, 64]> k_3_cast_fp16 = expand_dims(axes = k_3_axes_0, x = var_205_cast_fp16)[name = string("k_3_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_1_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22351424))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23154304))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_1_v_bias_to_fp16 = const()[name = string("trunk_layers_1_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23156160)))];
            tensor<fp16, [130, 896]> linear_9_cast_fp16 = linear(bias = trunk_layers_1_v_bias_to_fp16, weight = trunk_layers_1_v_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [3]> var_210 = const()[name = string("op_210"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_211_cast_fp16 = reshape(shape = var_210, x = linear_9_cast_fp16)[name = string("op_211_cast_fp16")];
            tensor<int32, [3]> var_212_perm_0 = const()[name = string("op_212_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_3_axes_0 = const()[name = string("v_3_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_212_cast_fp16 = transpose(perm = var_212_perm_0, x = var_211_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 14, 130, 64]> v_3_cast_fp16 = expand_dims(axes = v_3_axes_0, x = var_212_cast_fp16)[name = string("v_3_cast_fp16")];
            fp16 mul_1_y_0_to_fp16 = const()[name = string("mul_1_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_1_cast_fp16 = mul(x = q_3_cast_fp16, y = mul_1_y_0_to_fp16)[name = string("mul_1_cast_fp16")];
            bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(true)];
            bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_1_cast_fp16, y = k_3_cast_fp16)[name = string("matmul_1_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = attn_mask_to_fp16)[name = string("add_1_cast_fp16")];
            int32 softmax_1_axis_0 = const()[name = string("softmax_1_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = string("softmax_1_cast_fp16")];
            bool a_3_transpose_x_0 = const()[name = string("a_3_transpose_x_0"), val = bool(false)];
            bool a_3_transpose_y_0 = const()[name = string("a_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_3_cast_fp16 = matmul(transpose_x = a_3_transpose_x_0, transpose_y = a_3_transpose_y_0, x = softmax_1_cast_fp16, y = v_3_cast_fp16)[name = string("a_3_cast_fp16")];
            tensor<int32, [1]> var_215_axes_0 = const()[name = string("op_215_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_215_cast_fp16 = squeeze(axes = var_215_axes_0, x = a_3_cast_fp16)[name = string("op_215_cast_fp16")];
            tensor<int32, [3]> var_216_perm_0 = const()[name = string("op_216_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_217 = const()[name = string("op_217"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_216_cast_fp16 = transpose(perm = var_216_perm_0, x = var_215_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [130, 896]> input_25_cast_fp16 = reshape(shape = var_217, x = var_216_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_1_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23158016))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23960896))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_1_out_bias_to_fp16 = const()[name = string("trunk_layers_1_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23962752)))];
            tensor<fp16, [130, 896]> linear_10_cast_fp16 = linear(bias = trunk_layers_1_out_bias_to_fp16, weight = trunk_layers_1_out_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [130, 896]> input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_10_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<int32, [1]> input_29_axes_0 = const()[name = string("input_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_1_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_1_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23964608)))];
            tensor<fp16, [896]> trunk_layers_1_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_1_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23966464)))];
            tensor<fp16, [130, 896]> input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = trunk_layers_1_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_1_mlp_ln_weight_to_fp16, x = input_27_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_1_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23968320))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27179648))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_1_fc1_bias_to_fp16 = const()[name = string("trunk_layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27186880)))];
            tensor<fp16, [130, 3584]> linear_11_cast_fp16 = linear(bias = trunk_layers_1_fc1_bias_to_fp16, weight = trunk_layers_1_fc1_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_11_cast_fp16")];
            string input_31_mode_0 = const()[name = string("input_31_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = linear_11_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_1_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27194112))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30405440))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_1_fc2_bias_to_fp16 = const()[name = string("trunk_layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30407296)))];
            tensor<fp16, [130, 896]> linear_12_cast_fp16 = linear(bias = trunk_layers_1_fc2_bias_to_fp16, weight = trunk_layers_1_fc2_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [130, 896]> input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_12_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<int32, [1]> input_35_axes_0 = const()[name = string("input_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_2_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_2_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30409152)))];
            tensor<fp16, [896]> trunk_layers_2_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_2_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30411008)))];
            tensor<fp16, [130, 896]> input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = trunk_layers_2_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_2_attn_ln_weight_to_fp16, x = input_33_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_2_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30412864))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31215744))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_2_q_bias_to_fp16 = const()[name = string("trunk_layers_2_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31217600)))];
            tensor<fp16, [130, 896]> linear_13_cast_fp16 = linear(bias = trunk_layers_2_q_bias_to_fp16, weight = trunk_layers_2_q_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<int32, [3]> var_251 = const()[name = string("op_251"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_252_cast_fp16 = reshape(shape = var_251, x = linear_13_cast_fp16)[name = string("op_252_cast_fp16")];
            tensor<int32, [3]> var_253_perm_0 = const()[name = string("op_253_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_5_axes_0 = const()[name = string("q_5_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_253_cast_fp16 = transpose(perm = var_253_perm_0, x = var_252_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 14, 130, 64]> q_5_cast_fp16 = expand_dims(axes = q_5_axes_0, x = var_253_cast_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_2_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31219456))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32022336))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_2_k_bias_to_fp16 = const()[name = string("trunk_layers_2_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32024192)))];
            tensor<fp16, [130, 896]> linear_14_cast_fp16 = linear(bias = trunk_layers_2_k_bias_to_fp16, weight = trunk_layers_2_k_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [3]> var_258 = const()[name = string("op_258"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_259_cast_fp16 = reshape(shape = var_258, x = linear_14_cast_fp16)[name = string("op_259_cast_fp16")];
            tensor<int32, [3]> var_260_perm_0 = const()[name = string("op_260_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_5_axes_0 = const()[name = string("k_5_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_260_cast_fp16 = transpose(perm = var_260_perm_0, x = var_259_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 14, 130, 64]> k_5_cast_fp16 = expand_dims(axes = k_5_axes_0, x = var_260_cast_fp16)[name = string("k_5_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_2_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32026048))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32828928))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_2_v_bias_to_fp16 = const()[name = string("trunk_layers_2_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32830784)))];
            tensor<fp16, [130, 896]> linear_15_cast_fp16 = linear(bias = trunk_layers_2_v_bias_to_fp16, weight = trunk_layers_2_v_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [3]> var_265 = const()[name = string("op_265"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_266_cast_fp16 = reshape(shape = var_265, x = linear_15_cast_fp16)[name = string("op_266_cast_fp16")];
            tensor<int32, [3]> var_267_perm_0 = const()[name = string("op_267_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_5_axes_0 = const()[name = string("v_5_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_267_cast_fp16 = transpose(perm = var_267_perm_0, x = var_266_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 14, 130, 64]> v_5_cast_fp16 = expand_dims(axes = v_5_axes_0, x = var_267_cast_fp16)[name = string("v_5_cast_fp16")];
            fp16 mul_2_y_0_to_fp16 = const()[name = string("mul_2_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_2_cast_fp16 = mul(x = q_5_cast_fp16, y = mul_2_y_0_to_fp16)[name = string("mul_2_cast_fp16")];
            bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(true)];
            bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_2_cast_fp16, y = k_5_cast_fp16)[name = string("matmul_2_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = attn_mask_to_fp16)[name = string("add_2_cast_fp16")];
            int32 softmax_2_axis_0 = const()[name = string("softmax_2_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16)[name = string("softmax_2_cast_fp16")];
            bool a_5_transpose_x_0 = const()[name = string("a_5_transpose_x_0"), val = bool(false)];
            bool a_5_transpose_y_0 = const()[name = string("a_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_5_cast_fp16 = matmul(transpose_x = a_5_transpose_x_0, transpose_y = a_5_transpose_y_0, x = softmax_2_cast_fp16, y = v_5_cast_fp16)[name = string("a_5_cast_fp16")];
            tensor<int32, [1]> var_270_axes_0 = const()[name = string("op_270_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_270_cast_fp16 = squeeze(axes = var_270_axes_0, x = a_5_cast_fp16)[name = string("op_270_cast_fp16")];
            tensor<int32, [3]> var_271_perm_0 = const()[name = string("op_271_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_272 = const()[name = string("op_272"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_271_cast_fp16 = transpose(perm = var_271_perm_0, x = var_270_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [130, 896]> input_37_cast_fp16 = reshape(shape = var_272, x = var_271_cast_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_2_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32832640))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33635520))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_2_out_bias_to_fp16 = const()[name = string("trunk_layers_2_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33637376)))];
            tensor<fp16, [130, 896]> linear_16_cast_fp16 = linear(bias = trunk_layers_2_out_bias_to_fp16, weight = trunk_layers_2_out_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<fp16, [130, 896]> input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_16_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<int32, [1]> input_41_axes_0 = const()[name = string("input_41_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_2_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_2_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33639232)))];
            tensor<fp16, [896]> trunk_layers_2_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_2_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33641088)))];
            tensor<fp16, [130, 896]> input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = trunk_layers_2_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_2_mlp_ln_weight_to_fp16, x = input_39_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_2_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33642944))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36854272))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_2_fc1_bias_to_fp16 = const()[name = string("trunk_layers_2_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36861504)))];
            tensor<fp16, [130, 3584]> linear_17_cast_fp16 = linear(bias = trunk_layers_2_fc1_bias_to_fp16, weight = trunk_layers_2_fc1_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_17_cast_fp16")];
            string input_43_mode_0 = const()[name = string("input_43_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_43_cast_fp16 = gelu(mode = input_43_mode_0, x = linear_17_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_2_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36868736))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40080064))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_2_fc2_bias_to_fp16 = const()[name = string("trunk_layers_2_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40081920)))];
            tensor<fp16, [130, 896]> linear_18_cast_fp16 = linear(bias = trunk_layers_2_fc2_bias_to_fp16, weight = trunk_layers_2_fc2_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [130, 896]> input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_18_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<int32, [1]> input_47_axes_0 = const()[name = string("input_47_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_3_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_3_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40083776)))];
            tensor<fp16, [896]> trunk_layers_3_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_3_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40085632)))];
            tensor<fp16, [130, 896]> input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = trunk_layers_3_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_3_attn_ln_weight_to_fp16, x = input_45_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_3_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40087488))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40890368))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_3_q_bias_to_fp16 = const()[name = string("trunk_layers_3_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40892224)))];
            tensor<fp16, [130, 896]> linear_19_cast_fp16 = linear(bias = trunk_layers_3_q_bias_to_fp16, weight = trunk_layers_3_q_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<int32, [3]> var_306 = const()[name = string("op_306"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_307_cast_fp16 = reshape(shape = var_306, x = linear_19_cast_fp16)[name = string("op_307_cast_fp16")];
            tensor<int32, [3]> var_308_perm_0 = const()[name = string("op_308_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_7_axes_0 = const()[name = string("q_7_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_308_cast_fp16 = transpose(perm = var_308_perm_0, x = var_307_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 14, 130, 64]> q_7_cast_fp16 = expand_dims(axes = q_7_axes_0, x = var_308_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_3_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40894080))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41696960))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_3_k_bias_to_fp16 = const()[name = string("trunk_layers_3_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41698816)))];
            tensor<fp16, [130, 896]> linear_20_cast_fp16 = linear(bias = trunk_layers_3_k_bias_to_fp16, weight = trunk_layers_3_k_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [3]> var_313 = const()[name = string("op_313"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_314_cast_fp16 = reshape(shape = var_313, x = linear_20_cast_fp16)[name = string("op_314_cast_fp16")];
            tensor<int32, [3]> var_315_perm_0 = const()[name = string("op_315_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_7_axes_0 = const()[name = string("k_7_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_315_cast_fp16 = transpose(perm = var_315_perm_0, x = var_314_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 14, 130, 64]> k_7_cast_fp16 = expand_dims(axes = k_7_axes_0, x = var_315_cast_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_3_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41700672))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42503552))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_3_v_bias_to_fp16 = const()[name = string("trunk_layers_3_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42505408)))];
            tensor<fp16, [130, 896]> linear_21_cast_fp16 = linear(bias = trunk_layers_3_v_bias_to_fp16, weight = trunk_layers_3_v_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [3]> var_320 = const()[name = string("op_320"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_321_cast_fp16 = reshape(shape = var_320, x = linear_21_cast_fp16)[name = string("op_321_cast_fp16")];
            tensor<int32, [3]> var_322_perm_0 = const()[name = string("op_322_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_7_axes_0 = const()[name = string("v_7_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_322_cast_fp16 = transpose(perm = var_322_perm_0, x = var_321_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 14, 130, 64]> v_7_cast_fp16 = expand_dims(axes = v_7_axes_0, x = var_322_cast_fp16)[name = string("v_7_cast_fp16")];
            fp16 mul_3_y_0_to_fp16 = const()[name = string("mul_3_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_3_cast_fp16 = mul(x = q_7_cast_fp16, y = mul_3_y_0_to_fp16)[name = string("mul_3_cast_fp16")];
            bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(true)];
            bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_3_cast_fp16, y = k_7_cast_fp16)[name = string("matmul_3_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = attn_mask_to_fp16)[name = string("add_3_cast_fp16")];
            int32 softmax_3_axis_0 = const()[name = string("softmax_3_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = add_3_cast_fp16)[name = string("softmax_3_cast_fp16")];
            bool a_7_transpose_x_0 = const()[name = string("a_7_transpose_x_0"), val = bool(false)];
            bool a_7_transpose_y_0 = const()[name = string("a_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_7_cast_fp16 = matmul(transpose_x = a_7_transpose_x_0, transpose_y = a_7_transpose_y_0, x = softmax_3_cast_fp16, y = v_7_cast_fp16)[name = string("a_7_cast_fp16")];
            tensor<int32, [1]> var_325_axes_0 = const()[name = string("op_325_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_325_cast_fp16 = squeeze(axes = var_325_axes_0, x = a_7_cast_fp16)[name = string("op_325_cast_fp16")];
            tensor<int32, [3]> var_326_perm_0 = const()[name = string("op_326_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_327 = const()[name = string("op_327"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_326_cast_fp16 = transpose(perm = var_326_perm_0, x = var_325_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [130, 896]> input_49_cast_fp16 = reshape(shape = var_327, x = var_326_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_3_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42507264))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43310144))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_3_out_bias_to_fp16 = const()[name = string("trunk_layers_3_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43312000)))];
            tensor<fp16, [130, 896]> linear_22_cast_fp16 = linear(bias = trunk_layers_3_out_bias_to_fp16, weight = trunk_layers_3_out_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<fp16, [130, 896]> input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_22_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<int32, [1]> input_53_axes_0 = const()[name = string("input_53_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_3_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_3_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43313856)))];
            tensor<fp16, [896]> trunk_layers_3_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_3_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43315712)))];
            tensor<fp16, [130, 896]> input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = trunk_layers_3_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_3_mlp_ln_weight_to_fp16, x = input_51_cast_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_3_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43317568))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46528896))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_3_fc1_bias_to_fp16 = const()[name = string("trunk_layers_3_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46536128)))];
            tensor<fp16, [130, 3584]> linear_23_cast_fp16 = linear(bias = trunk_layers_3_fc1_bias_to_fp16, weight = trunk_layers_3_fc1_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_23_cast_fp16")];
            string input_55_mode_0 = const()[name = string("input_55_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_55_cast_fp16 = gelu(mode = input_55_mode_0, x = linear_23_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_3_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46543360))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49754688))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_3_fc2_bias_to_fp16 = const()[name = string("trunk_layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49756544)))];
            tensor<fp16, [130, 896]> linear_24_cast_fp16 = linear(bias = trunk_layers_3_fc2_bias_to_fp16, weight = trunk_layers_3_fc2_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [130, 896]> input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_24_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<int32, [1]> input_59_axes_0 = const()[name = string("input_59_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_4_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_4_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49758400)))];
            tensor<fp16, [896]> trunk_layers_4_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_4_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49760256)))];
            tensor<fp16, [130, 896]> input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = trunk_layers_4_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_4_attn_ln_weight_to_fp16, x = input_57_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_4_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49762112))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50564992))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_4_q_bias_to_fp16 = const()[name = string("trunk_layers_4_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50566848)))];
            tensor<fp16, [130, 896]> linear_25_cast_fp16 = linear(bias = trunk_layers_4_q_bias_to_fp16, weight = trunk_layers_4_q_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<int32, [3]> var_361 = const()[name = string("op_361"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_362_cast_fp16 = reshape(shape = var_361, x = linear_25_cast_fp16)[name = string("op_362_cast_fp16")];
            tensor<int32, [3]> var_363_perm_0 = const()[name = string("op_363_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_9_axes_0 = const()[name = string("q_9_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_363_cast_fp16 = transpose(perm = var_363_perm_0, x = var_362_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 14, 130, 64]> q_9_cast_fp16 = expand_dims(axes = q_9_axes_0, x = var_363_cast_fp16)[name = string("q_9_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_4_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50568704))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51371584))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_4_k_bias_to_fp16 = const()[name = string("trunk_layers_4_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51373440)))];
            tensor<fp16, [130, 896]> linear_26_cast_fp16 = linear(bias = trunk_layers_4_k_bias_to_fp16, weight = trunk_layers_4_k_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<int32, [3]> var_368 = const()[name = string("op_368"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_369_cast_fp16 = reshape(shape = var_368, x = linear_26_cast_fp16)[name = string("op_369_cast_fp16")];
            tensor<int32, [3]> var_370_perm_0 = const()[name = string("op_370_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_9_axes_0 = const()[name = string("k_9_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_370_cast_fp16 = transpose(perm = var_370_perm_0, x = var_369_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 14, 130, 64]> k_9_cast_fp16 = expand_dims(axes = k_9_axes_0, x = var_370_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_4_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51375296))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52178176))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_4_v_bias_to_fp16 = const()[name = string("trunk_layers_4_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52180032)))];
            tensor<fp16, [130, 896]> linear_27_cast_fp16 = linear(bias = trunk_layers_4_v_bias_to_fp16, weight = trunk_layers_4_v_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<int32, [3]> var_375 = const()[name = string("op_375"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_376_cast_fp16 = reshape(shape = var_375, x = linear_27_cast_fp16)[name = string("op_376_cast_fp16")];
            tensor<int32, [3]> var_377_perm_0 = const()[name = string("op_377_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_9_axes_0 = const()[name = string("v_9_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_377_cast_fp16 = transpose(perm = var_377_perm_0, x = var_376_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 14, 130, 64]> v_9_cast_fp16 = expand_dims(axes = v_9_axes_0, x = var_377_cast_fp16)[name = string("v_9_cast_fp16")];
            fp16 mul_4_y_0_to_fp16 = const()[name = string("mul_4_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_4_cast_fp16 = mul(x = q_9_cast_fp16, y = mul_4_y_0_to_fp16)[name = string("mul_4_cast_fp16")];
            bool matmul_4_transpose_y_0 = const()[name = string("matmul_4_transpose_y_0"), val = bool(true)];
            bool matmul_4_transpose_x_0 = const()[name = string("matmul_4_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_4_cast_fp16, y = k_9_cast_fp16)[name = string("matmul_4_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = attn_mask_to_fp16)[name = string("add_4_cast_fp16")];
            int32 softmax_4_axis_0 = const()[name = string("softmax_4_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = add_4_cast_fp16)[name = string("softmax_4_cast_fp16")];
            bool a_9_transpose_x_0 = const()[name = string("a_9_transpose_x_0"), val = bool(false)];
            bool a_9_transpose_y_0 = const()[name = string("a_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_9_cast_fp16 = matmul(transpose_x = a_9_transpose_x_0, transpose_y = a_9_transpose_y_0, x = softmax_4_cast_fp16, y = v_9_cast_fp16)[name = string("a_9_cast_fp16")];
            tensor<int32, [1]> var_380_axes_0 = const()[name = string("op_380_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_380_cast_fp16 = squeeze(axes = var_380_axes_0, x = a_9_cast_fp16)[name = string("op_380_cast_fp16")];
            tensor<int32, [3]> var_381_perm_0 = const()[name = string("op_381_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_382 = const()[name = string("op_382"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_381_cast_fp16 = transpose(perm = var_381_perm_0, x = var_380_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [130, 896]> input_61_cast_fp16 = reshape(shape = var_382, x = var_381_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_4_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52181888))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52984768))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_4_out_bias_to_fp16 = const()[name = string("trunk_layers_4_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52986624)))];
            tensor<fp16, [130, 896]> linear_28_cast_fp16 = linear(bias = trunk_layers_4_out_bias_to_fp16, weight = trunk_layers_4_out_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<fp16, [130, 896]> input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_28_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<int32, [1]> input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_4_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_4_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52988480)))];
            tensor<fp16, [896]> trunk_layers_4_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_4_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52990336)))];
            tensor<fp16, [130, 896]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = trunk_layers_4_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_4_mlp_ln_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_4_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52992192))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56203520))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_4_fc1_bias_to_fp16 = const()[name = string("trunk_layers_4_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56210752)))];
            tensor<fp16, [130, 3584]> linear_29_cast_fp16 = linear(bias = trunk_layers_4_fc1_bias_to_fp16, weight = trunk_layers_4_fc1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_29_cast_fp16")];
            string input_67_mode_0 = const()[name = string("input_67_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_67_cast_fp16 = gelu(mode = input_67_mode_0, x = linear_29_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_4_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56217984))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59429312))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_4_fc2_bias_to_fp16 = const()[name = string("trunk_layers_4_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59431168)))];
            tensor<fp16, [130, 896]> linear_30_cast_fp16 = linear(bias = trunk_layers_4_fc2_bias_to_fp16, weight = trunk_layers_4_fc2_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<fp16, [130, 896]> input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_30_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<int32, [1]> input_71_axes_0 = const()[name = string("input_71_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_5_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_5_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59433024)))];
            tensor<fp16, [896]> trunk_layers_5_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_5_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59434880)))];
            tensor<fp16, [130, 896]> input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = trunk_layers_5_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_5_attn_ln_weight_to_fp16, x = input_69_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_5_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59436736))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60239616))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_5_q_bias_to_fp16 = const()[name = string("trunk_layers_5_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60241472)))];
            tensor<fp16, [130, 896]> linear_31_cast_fp16 = linear(bias = trunk_layers_5_q_bias_to_fp16, weight = trunk_layers_5_q_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<int32, [3]> var_416 = const()[name = string("op_416"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_417_cast_fp16 = reshape(shape = var_416, x = linear_31_cast_fp16)[name = string("op_417_cast_fp16")];
            tensor<int32, [3]> var_418_perm_0 = const()[name = string("op_418_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_11_axes_0 = const()[name = string("q_11_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_418_cast_fp16 = transpose(perm = var_418_perm_0, x = var_417_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 14, 130, 64]> q_11_cast_fp16 = expand_dims(axes = q_11_axes_0, x = var_418_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_5_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60243328))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61046208))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_5_k_bias_to_fp16 = const()[name = string("trunk_layers_5_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61048064)))];
            tensor<fp16, [130, 896]> linear_32_cast_fp16 = linear(bias = trunk_layers_5_k_bias_to_fp16, weight = trunk_layers_5_k_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<int32, [3]> var_423 = const()[name = string("op_423"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_424_cast_fp16 = reshape(shape = var_423, x = linear_32_cast_fp16)[name = string("op_424_cast_fp16")];
            tensor<int32, [3]> var_425_perm_0 = const()[name = string("op_425_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_11_axes_0 = const()[name = string("k_11_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_425_cast_fp16 = transpose(perm = var_425_perm_0, x = var_424_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 14, 130, 64]> k_11_cast_fp16 = expand_dims(axes = k_11_axes_0, x = var_425_cast_fp16)[name = string("k_11_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_5_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61049920))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61852800))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_5_v_bias_to_fp16 = const()[name = string("trunk_layers_5_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61854656)))];
            tensor<fp16, [130, 896]> linear_33_cast_fp16 = linear(bias = trunk_layers_5_v_bias_to_fp16, weight = trunk_layers_5_v_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<int32, [3]> var_430 = const()[name = string("op_430"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_431_cast_fp16 = reshape(shape = var_430, x = linear_33_cast_fp16)[name = string("op_431_cast_fp16")];
            tensor<int32, [3]> var_432_perm_0 = const()[name = string("op_432_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_11_axes_0 = const()[name = string("v_11_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_432_cast_fp16 = transpose(perm = var_432_perm_0, x = var_431_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 14, 130, 64]> v_11_cast_fp16 = expand_dims(axes = v_11_axes_0, x = var_432_cast_fp16)[name = string("v_11_cast_fp16")];
            fp16 mul_5_y_0_to_fp16 = const()[name = string("mul_5_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_5_cast_fp16 = mul(x = q_11_cast_fp16, y = mul_5_y_0_to_fp16)[name = string("mul_5_cast_fp16")];
            bool matmul_5_transpose_y_0 = const()[name = string("matmul_5_transpose_y_0"), val = bool(true)];
            bool matmul_5_transpose_x_0 = const()[name = string("matmul_5_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_5_cast_fp16, y = k_11_cast_fp16)[name = string("matmul_5_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = attn_mask_to_fp16)[name = string("add_5_cast_fp16")];
            int32 softmax_5_axis_0 = const()[name = string("softmax_5_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = add_5_cast_fp16)[name = string("softmax_5_cast_fp16")];
            bool a_11_transpose_x_0 = const()[name = string("a_11_transpose_x_0"), val = bool(false)];
            bool a_11_transpose_y_0 = const()[name = string("a_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_11_cast_fp16 = matmul(transpose_x = a_11_transpose_x_0, transpose_y = a_11_transpose_y_0, x = softmax_5_cast_fp16, y = v_11_cast_fp16)[name = string("a_11_cast_fp16")];
            tensor<int32, [1]> var_435_axes_0 = const()[name = string("op_435_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_435_cast_fp16 = squeeze(axes = var_435_axes_0, x = a_11_cast_fp16)[name = string("op_435_cast_fp16")];
            tensor<int32, [3]> var_436_perm_0 = const()[name = string("op_436_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_437 = const()[name = string("op_437"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_436_cast_fp16 = transpose(perm = var_436_perm_0, x = var_435_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [130, 896]> input_73_cast_fp16 = reshape(shape = var_437, x = var_436_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_5_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61856512))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62659392))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_5_out_bias_to_fp16 = const()[name = string("trunk_layers_5_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62661248)))];
            tensor<fp16, [130, 896]> linear_34_cast_fp16 = linear(bias = trunk_layers_5_out_bias_to_fp16, weight = trunk_layers_5_out_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [130, 896]> input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_34_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<int32, [1]> input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_5_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_5_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62663104)))];
            tensor<fp16, [896]> trunk_layers_5_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_5_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62664960)))];
            tensor<fp16, [130, 896]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = trunk_layers_5_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_5_mlp_ln_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_5_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62666816))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65878144))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_5_fc1_bias_to_fp16 = const()[name = string("trunk_layers_5_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65885376)))];
            tensor<fp16, [130, 3584]> linear_35_cast_fp16 = linear(bias = trunk_layers_5_fc1_bias_to_fp16, weight = trunk_layers_5_fc1_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_35_cast_fp16")];
            string input_79_mode_0 = const()[name = string("input_79_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = linear_35_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_5_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65892608))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69103936))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_5_fc2_bias_to_fp16 = const()[name = string("trunk_layers_5_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69105792)))];
            tensor<fp16, [130, 896]> linear_36_cast_fp16 = linear(bias = trunk_layers_5_fc2_bias_to_fp16, weight = trunk_layers_5_fc2_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<fp16, [130, 896]> input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_36_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<int32, [1]> input_83_axes_0 = const()[name = string("input_83_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_6_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_6_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69107648)))];
            tensor<fp16, [896]> trunk_layers_6_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_6_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69109504)))];
            tensor<fp16, [130, 896]> input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = trunk_layers_6_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_6_attn_ln_weight_to_fp16, x = input_81_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_6_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69111360))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69914240))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_6_q_bias_to_fp16 = const()[name = string("trunk_layers_6_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69916096)))];
            tensor<fp16, [130, 896]> linear_37_cast_fp16 = linear(bias = trunk_layers_6_q_bias_to_fp16, weight = trunk_layers_6_q_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [3]> var_471 = const()[name = string("op_471"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_472_cast_fp16 = reshape(shape = var_471, x = linear_37_cast_fp16)[name = string("op_472_cast_fp16")];
            tensor<int32, [3]> var_473_perm_0 = const()[name = string("op_473_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_13_axes_0 = const()[name = string("q_13_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_473_cast_fp16 = transpose(perm = var_473_perm_0, x = var_472_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 14, 130, 64]> q_13_cast_fp16 = expand_dims(axes = q_13_axes_0, x = var_473_cast_fp16)[name = string("q_13_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_6_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69917952))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70720832))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_6_k_bias_to_fp16 = const()[name = string("trunk_layers_6_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70722688)))];
            tensor<fp16, [130, 896]> linear_38_cast_fp16 = linear(bias = trunk_layers_6_k_bias_to_fp16, weight = trunk_layers_6_k_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [3]> var_478 = const()[name = string("op_478"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_479_cast_fp16 = reshape(shape = var_478, x = linear_38_cast_fp16)[name = string("op_479_cast_fp16")];
            tensor<int32, [3]> var_480_perm_0 = const()[name = string("op_480_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_13_axes_0 = const()[name = string("k_13_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_480_cast_fp16 = transpose(perm = var_480_perm_0, x = var_479_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 14, 130, 64]> k_13_cast_fp16 = expand_dims(axes = k_13_axes_0, x = var_480_cast_fp16)[name = string("k_13_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_6_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70724544))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71527424))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_6_v_bias_to_fp16 = const()[name = string("trunk_layers_6_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71529280)))];
            tensor<fp16, [130, 896]> linear_39_cast_fp16 = linear(bias = trunk_layers_6_v_bias_to_fp16, weight = trunk_layers_6_v_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<int32, [3]> var_485 = const()[name = string("op_485"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_486_cast_fp16 = reshape(shape = var_485, x = linear_39_cast_fp16)[name = string("op_486_cast_fp16")];
            tensor<int32, [3]> var_487_perm_0 = const()[name = string("op_487_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_13_axes_0 = const()[name = string("v_13_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_487_cast_fp16 = transpose(perm = var_487_perm_0, x = var_486_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 14, 130, 64]> v_13_cast_fp16 = expand_dims(axes = v_13_axes_0, x = var_487_cast_fp16)[name = string("v_13_cast_fp16")];
            fp16 mul_6_y_0_to_fp16 = const()[name = string("mul_6_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_6_cast_fp16 = mul(x = q_13_cast_fp16, y = mul_6_y_0_to_fp16)[name = string("mul_6_cast_fp16")];
            bool matmul_6_transpose_y_0 = const()[name = string("matmul_6_transpose_y_0"), val = bool(true)];
            bool matmul_6_transpose_x_0 = const()[name = string("matmul_6_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_6_cast_fp16, y = k_13_cast_fp16)[name = string("matmul_6_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_6_cast_fp16 = add(x = matmul_6_cast_fp16, y = attn_mask_to_fp16)[name = string("add_6_cast_fp16")];
            int32 softmax_6_axis_0 = const()[name = string("softmax_6_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = add_6_cast_fp16)[name = string("softmax_6_cast_fp16")];
            bool a_13_transpose_x_0 = const()[name = string("a_13_transpose_x_0"), val = bool(false)];
            bool a_13_transpose_y_0 = const()[name = string("a_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_13_cast_fp16 = matmul(transpose_x = a_13_transpose_x_0, transpose_y = a_13_transpose_y_0, x = softmax_6_cast_fp16, y = v_13_cast_fp16)[name = string("a_13_cast_fp16")];
            tensor<int32, [1]> var_490_axes_0 = const()[name = string("op_490_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_490_cast_fp16 = squeeze(axes = var_490_axes_0, x = a_13_cast_fp16)[name = string("op_490_cast_fp16")];
            tensor<int32, [3]> var_491_perm_0 = const()[name = string("op_491_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_492 = const()[name = string("op_492"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_491_cast_fp16 = transpose(perm = var_491_perm_0, x = var_490_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [130, 896]> input_85_cast_fp16 = reshape(shape = var_492, x = var_491_cast_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_6_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71531136))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72334016))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_6_out_bias_to_fp16 = const()[name = string("trunk_layers_6_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72335872)))];
            tensor<fp16, [130, 896]> linear_40_cast_fp16 = linear(bias = trunk_layers_6_out_bias_to_fp16, weight = trunk_layers_6_out_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [130, 896]> input_87_cast_fp16 = add(x = input_81_cast_fp16, y = linear_40_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<int32, [1]> input_89_axes_0 = const()[name = string("input_89_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_6_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_6_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72337728)))];
            tensor<fp16, [896]> trunk_layers_6_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_6_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72339584)))];
            tensor<fp16, [130, 896]> input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = trunk_layers_6_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_6_mlp_ln_weight_to_fp16, x = input_87_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_6_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72341440))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75552768))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_6_fc1_bias_to_fp16 = const()[name = string("trunk_layers_6_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75560000)))];
            tensor<fp16, [130, 3584]> linear_41_cast_fp16 = linear(bias = trunk_layers_6_fc1_bias_to_fp16, weight = trunk_layers_6_fc1_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_41_cast_fp16")];
            string input_91_mode_0 = const()[name = string("input_91_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = linear_41_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_6_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75567232))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78778560))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_6_fc2_bias_to_fp16 = const()[name = string("trunk_layers_6_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78780416)))];
            tensor<fp16, [130, 896]> linear_42_cast_fp16 = linear(bias = trunk_layers_6_fc2_bias_to_fp16, weight = trunk_layers_6_fc2_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [130, 896]> input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_42_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<int32, [1]> input_95_axes_0 = const()[name = string("input_95_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_7_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_7_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78782272)))];
            tensor<fp16, [896]> trunk_layers_7_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_7_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78784128)))];
            tensor<fp16, [130, 896]> input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = trunk_layers_7_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_7_attn_ln_weight_to_fp16, x = input_93_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_7_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78785984))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79588864))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_7_q_bias_to_fp16 = const()[name = string("trunk_layers_7_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79590720)))];
            tensor<fp16, [130, 896]> linear_43_cast_fp16 = linear(bias = trunk_layers_7_q_bias_to_fp16, weight = trunk_layers_7_q_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [3]> var_526 = const()[name = string("op_526"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_527_cast_fp16 = reshape(shape = var_526, x = linear_43_cast_fp16)[name = string("op_527_cast_fp16")];
            tensor<int32, [3]> var_528_perm_0 = const()[name = string("op_528_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_15_axes_0 = const()[name = string("q_15_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_528_cast_fp16 = transpose(perm = var_528_perm_0, x = var_527_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 14, 130, 64]> q_15_cast_fp16 = expand_dims(axes = q_15_axes_0, x = var_528_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_7_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79592576))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80395456))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_7_k_bias_to_fp16 = const()[name = string("trunk_layers_7_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80397312)))];
            tensor<fp16, [130, 896]> linear_44_cast_fp16 = linear(bias = trunk_layers_7_k_bias_to_fp16, weight = trunk_layers_7_k_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [3]> var_533 = const()[name = string("op_533"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_534_cast_fp16 = reshape(shape = var_533, x = linear_44_cast_fp16)[name = string("op_534_cast_fp16")];
            tensor<int32, [3]> var_535_perm_0 = const()[name = string("op_535_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_15_axes_0 = const()[name = string("k_15_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_535_cast_fp16 = transpose(perm = var_535_perm_0, x = var_534_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 14, 130, 64]> k_15_cast_fp16 = expand_dims(axes = k_15_axes_0, x = var_535_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_7_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80399168))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81202048))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_7_v_bias_to_fp16 = const()[name = string("trunk_layers_7_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81203904)))];
            tensor<fp16, [130, 896]> linear_45_cast_fp16 = linear(bias = trunk_layers_7_v_bias_to_fp16, weight = trunk_layers_7_v_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<int32, [3]> var_540 = const()[name = string("op_540"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_541_cast_fp16 = reshape(shape = var_540, x = linear_45_cast_fp16)[name = string("op_541_cast_fp16")];
            tensor<int32, [3]> var_542_perm_0 = const()[name = string("op_542_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_15_axes_0 = const()[name = string("v_15_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_542_cast_fp16 = transpose(perm = var_542_perm_0, x = var_541_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 14, 130, 64]> v_15_cast_fp16 = expand_dims(axes = v_15_axes_0, x = var_542_cast_fp16)[name = string("v_15_cast_fp16")];
            fp16 mul_7_y_0_to_fp16 = const()[name = string("mul_7_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_7_cast_fp16 = mul(x = q_15_cast_fp16, y = mul_7_y_0_to_fp16)[name = string("mul_7_cast_fp16")];
            bool matmul_7_transpose_y_0 = const()[name = string("matmul_7_transpose_y_0"), val = bool(true)];
            bool matmul_7_transpose_x_0 = const()[name = string("matmul_7_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_7_cast_fp16, y = k_15_cast_fp16)[name = string("matmul_7_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_7_cast_fp16 = add(x = matmul_7_cast_fp16, y = attn_mask_to_fp16)[name = string("add_7_cast_fp16")];
            int32 softmax_7_axis_0 = const()[name = string("softmax_7_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = add_7_cast_fp16)[name = string("softmax_7_cast_fp16")];
            bool a_15_transpose_x_0 = const()[name = string("a_15_transpose_x_0"), val = bool(false)];
            bool a_15_transpose_y_0 = const()[name = string("a_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_15_cast_fp16 = matmul(transpose_x = a_15_transpose_x_0, transpose_y = a_15_transpose_y_0, x = softmax_7_cast_fp16, y = v_15_cast_fp16)[name = string("a_15_cast_fp16")];
            tensor<int32, [1]> var_545_axes_0 = const()[name = string("op_545_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_545_cast_fp16 = squeeze(axes = var_545_axes_0, x = a_15_cast_fp16)[name = string("op_545_cast_fp16")];
            tensor<int32, [3]> var_546_perm_0 = const()[name = string("op_546_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_547 = const()[name = string("op_547"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_546_cast_fp16 = transpose(perm = var_546_perm_0, x = var_545_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [130, 896]> input_97_cast_fp16 = reshape(shape = var_547, x = var_546_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_7_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81205760))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82008640))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_7_out_bias_to_fp16 = const()[name = string("trunk_layers_7_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82010496)))];
            tensor<fp16, [130, 896]> linear_46_cast_fp16 = linear(bias = trunk_layers_7_out_bias_to_fp16, weight = trunk_layers_7_out_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [130, 896]> input_99_cast_fp16 = add(x = input_93_cast_fp16, y = linear_46_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<int32, [1]> input_101_axes_0 = const()[name = string("input_101_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_7_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_7_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82012352)))];
            tensor<fp16, [896]> trunk_layers_7_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_7_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82014208)))];
            tensor<fp16, [130, 896]> input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = trunk_layers_7_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_7_mlp_ln_weight_to_fp16, x = input_99_cast_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_7_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82016064))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85227392))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_7_fc1_bias_to_fp16 = const()[name = string("trunk_layers_7_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85234624)))];
            tensor<fp16, [130, 3584]> linear_47_cast_fp16 = linear(bias = trunk_layers_7_fc1_bias_to_fp16, weight = trunk_layers_7_fc1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_47_cast_fp16")];
            string input_103_mode_0 = const()[name = string("input_103_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_103_cast_fp16 = gelu(mode = input_103_mode_0, x = linear_47_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_7_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85241856))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88453184))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_7_fc2_bias_to_fp16 = const()[name = string("trunk_layers_7_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88455040)))];
            tensor<fp16, [130, 896]> linear_48_cast_fp16 = linear(bias = trunk_layers_7_fc2_bias_to_fp16, weight = trunk_layers_7_fc2_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [130, 896]> input_105_cast_fp16 = add(x = input_99_cast_fp16, y = linear_48_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<int32, [1]> input_107_axes_0 = const()[name = string("input_107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_8_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_8_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88456896)))];
            tensor<fp16, [896]> trunk_layers_8_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_8_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88458752)))];
            tensor<fp16, [130, 896]> input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = trunk_layers_8_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_8_attn_ln_weight_to_fp16, x = input_105_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_8_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88460608))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89263488))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_8_q_bias_to_fp16 = const()[name = string("trunk_layers_8_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89265344)))];
            tensor<fp16, [130, 896]> linear_49_cast_fp16 = linear(bias = trunk_layers_8_q_bias_to_fp16, weight = trunk_layers_8_q_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [3]> var_581 = const()[name = string("op_581"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_582_cast_fp16 = reshape(shape = var_581, x = linear_49_cast_fp16)[name = string("op_582_cast_fp16")];
            tensor<int32, [3]> var_583_perm_0 = const()[name = string("op_583_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_17_axes_0 = const()[name = string("q_17_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_583_cast_fp16 = transpose(perm = var_583_perm_0, x = var_582_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 14, 130, 64]> q_17_cast_fp16 = expand_dims(axes = q_17_axes_0, x = var_583_cast_fp16)[name = string("q_17_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_8_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89267200))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90070080))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_8_k_bias_to_fp16 = const()[name = string("trunk_layers_8_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90071936)))];
            tensor<fp16, [130, 896]> linear_50_cast_fp16 = linear(bias = trunk_layers_8_k_bias_to_fp16, weight = trunk_layers_8_k_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [3]> var_588 = const()[name = string("op_588"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_589_cast_fp16 = reshape(shape = var_588, x = linear_50_cast_fp16)[name = string("op_589_cast_fp16")];
            tensor<int32, [3]> var_590_perm_0 = const()[name = string("op_590_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_17_axes_0 = const()[name = string("k_17_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_590_cast_fp16 = transpose(perm = var_590_perm_0, x = var_589_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 14, 130, 64]> k_17_cast_fp16 = expand_dims(axes = k_17_axes_0, x = var_590_cast_fp16)[name = string("k_17_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_8_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90073792))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90876672))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_8_v_bias_to_fp16 = const()[name = string("trunk_layers_8_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90878528)))];
            tensor<fp16, [130, 896]> linear_51_cast_fp16 = linear(bias = trunk_layers_8_v_bias_to_fp16, weight = trunk_layers_8_v_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [3]> var_595 = const()[name = string("op_595"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_596_cast_fp16 = reshape(shape = var_595, x = linear_51_cast_fp16)[name = string("op_596_cast_fp16")];
            tensor<int32, [3]> var_597_perm_0 = const()[name = string("op_597_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_17_axes_0 = const()[name = string("v_17_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_597_cast_fp16 = transpose(perm = var_597_perm_0, x = var_596_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 14, 130, 64]> v_17_cast_fp16 = expand_dims(axes = v_17_axes_0, x = var_597_cast_fp16)[name = string("v_17_cast_fp16")];
            fp16 mul_8_y_0_to_fp16 = const()[name = string("mul_8_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_8_cast_fp16 = mul(x = q_17_cast_fp16, y = mul_8_y_0_to_fp16)[name = string("mul_8_cast_fp16")];
            bool matmul_8_transpose_y_0 = const()[name = string("matmul_8_transpose_y_0"), val = bool(true)];
            bool matmul_8_transpose_x_0 = const()[name = string("matmul_8_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_8_cast_fp16, y = k_17_cast_fp16)[name = string("matmul_8_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_8_cast_fp16 = add(x = matmul_8_cast_fp16, y = attn_mask_to_fp16)[name = string("add_8_cast_fp16")];
            int32 softmax_8_axis_0 = const()[name = string("softmax_8_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = add_8_cast_fp16)[name = string("softmax_8_cast_fp16")];
            bool a_17_transpose_x_0 = const()[name = string("a_17_transpose_x_0"), val = bool(false)];
            bool a_17_transpose_y_0 = const()[name = string("a_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_17_cast_fp16 = matmul(transpose_x = a_17_transpose_x_0, transpose_y = a_17_transpose_y_0, x = softmax_8_cast_fp16, y = v_17_cast_fp16)[name = string("a_17_cast_fp16")];
            tensor<int32, [1]> var_600_axes_0 = const()[name = string("op_600_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_600_cast_fp16 = squeeze(axes = var_600_axes_0, x = a_17_cast_fp16)[name = string("op_600_cast_fp16")];
            tensor<int32, [3]> var_601_perm_0 = const()[name = string("op_601_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_602 = const()[name = string("op_602"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_601_cast_fp16 = transpose(perm = var_601_perm_0, x = var_600_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [130, 896]> input_109_cast_fp16 = reshape(shape = var_602, x = var_601_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_8_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90880384))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91683264))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_8_out_bias_to_fp16 = const()[name = string("trunk_layers_8_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91685120)))];
            tensor<fp16, [130, 896]> linear_52_cast_fp16 = linear(bias = trunk_layers_8_out_bias_to_fp16, weight = trunk_layers_8_out_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [130, 896]> input_111_cast_fp16 = add(x = input_105_cast_fp16, y = linear_52_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<int32, [1]> input_113_axes_0 = const()[name = string("input_113_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_8_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_8_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91686976)))];
            tensor<fp16, [896]> trunk_layers_8_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_8_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91688832)))];
            tensor<fp16, [130, 896]> input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = trunk_layers_8_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_8_mlp_ln_weight_to_fp16, x = input_111_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_8_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91690688))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94902016))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_8_fc1_bias_to_fp16 = const()[name = string("trunk_layers_8_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94909248)))];
            tensor<fp16, [130, 3584]> linear_53_cast_fp16 = linear(bias = trunk_layers_8_fc1_bias_to_fp16, weight = trunk_layers_8_fc1_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_53_cast_fp16")];
            string input_115_mode_0 = const()[name = string("input_115_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_115_cast_fp16 = gelu(mode = input_115_mode_0, x = linear_53_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_8_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94916480))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98127808))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_8_fc2_bias_to_fp16 = const()[name = string("trunk_layers_8_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98129664)))];
            tensor<fp16, [130, 896]> linear_54_cast_fp16 = linear(bias = trunk_layers_8_fc2_bias_to_fp16, weight = trunk_layers_8_fc2_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [130, 896]> input_117_cast_fp16 = add(x = input_111_cast_fp16, y = linear_54_cast_fp16)[name = string("input_117_cast_fp16")];
            tensor<int32, [1]> input_119_axes_0 = const()[name = string("input_119_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_9_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_9_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98131520)))];
            tensor<fp16, [896]> trunk_layers_9_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_9_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98133376)))];
            tensor<fp16, [130, 896]> input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = trunk_layers_9_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_9_attn_ln_weight_to_fp16, x = input_117_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_9_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98135232))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98938112))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_9_q_bias_to_fp16 = const()[name = string("trunk_layers_9_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98939968)))];
            tensor<fp16, [130, 896]> linear_55_cast_fp16 = linear(bias = trunk_layers_9_q_bias_to_fp16, weight = trunk_layers_9_q_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<int32, [3]> var_636 = const()[name = string("op_636"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_637_cast_fp16 = reshape(shape = var_636, x = linear_55_cast_fp16)[name = string("op_637_cast_fp16")];
            tensor<int32, [3]> var_638_perm_0 = const()[name = string("op_638_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_19_axes_0 = const()[name = string("q_19_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_638_cast_fp16 = transpose(perm = var_638_perm_0, x = var_637_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 14, 130, 64]> q_19_cast_fp16 = expand_dims(axes = q_19_axes_0, x = var_638_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_9_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98941824))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99744704))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_9_k_bias_to_fp16 = const()[name = string("trunk_layers_9_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99746560)))];
            tensor<fp16, [130, 896]> linear_56_cast_fp16 = linear(bias = trunk_layers_9_k_bias_to_fp16, weight = trunk_layers_9_k_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [3]> var_643 = const()[name = string("op_643"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_644_cast_fp16 = reshape(shape = var_643, x = linear_56_cast_fp16)[name = string("op_644_cast_fp16")];
            tensor<int32, [3]> var_645_perm_0 = const()[name = string("op_645_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_19_axes_0 = const()[name = string("k_19_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_645_cast_fp16 = transpose(perm = var_645_perm_0, x = var_644_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 14, 130, 64]> k_19_cast_fp16 = expand_dims(axes = k_19_axes_0, x = var_645_cast_fp16)[name = string("k_19_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_9_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99748416))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100551296))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_9_v_bias_to_fp16 = const()[name = string("trunk_layers_9_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100553152)))];
            tensor<fp16, [130, 896]> linear_57_cast_fp16 = linear(bias = trunk_layers_9_v_bias_to_fp16, weight = trunk_layers_9_v_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [3]> var_650 = const()[name = string("op_650"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_651_cast_fp16 = reshape(shape = var_650, x = linear_57_cast_fp16)[name = string("op_651_cast_fp16")];
            tensor<int32, [3]> var_652_perm_0 = const()[name = string("op_652_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_19_axes_0 = const()[name = string("v_19_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_652_cast_fp16 = transpose(perm = var_652_perm_0, x = var_651_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 14, 130, 64]> v_19_cast_fp16 = expand_dims(axes = v_19_axes_0, x = var_652_cast_fp16)[name = string("v_19_cast_fp16")];
            fp16 mul_9_y_0_to_fp16 = const()[name = string("mul_9_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_9_cast_fp16 = mul(x = q_19_cast_fp16, y = mul_9_y_0_to_fp16)[name = string("mul_9_cast_fp16")];
            bool matmul_9_transpose_y_0 = const()[name = string("matmul_9_transpose_y_0"), val = bool(true)];
            bool matmul_9_transpose_x_0 = const()[name = string("matmul_9_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_9_cast_fp16, y = k_19_cast_fp16)[name = string("matmul_9_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_9_cast_fp16 = add(x = matmul_9_cast_fp16, y = attn_mask_to_fp16)[name = string("add_9_cast_fp16")];
            int32 softmax_9_axis_0 = const()[name = string("softmax_9_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = add_9_cast_fp16)[name = string("softmax_9_cast_fp16")];
            bool a_19_transpose_x_0 = const()[name = string("a_19_transpose_x_0"), val = bool(false)];
            bool a_19_transpose_y_0 = const()[name = string("a_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_19_cast_fp16 = matmul(transpose_x = a_19_transpose_x_0, transpose_y = a_19_transpose_y_0, x = softmax_9_cast_fp16, y = v_19_cast_fp16)[name = string("a_19_cast_fp16")];
            tensor<int32, [1]> var_655_axes_0 = const()[name = string("op_655_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_655_cast_fp16 = squeeze(axes = var_655_axes_0, x = a_19_cast_fp16)[name = string("op_655_cast_fp16")];
            tensor<int32, [3]> var_656_perm_0 = const()[name = string("op_656_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_657 = const()[name = string("op_657"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_656_cast_fp16 = transpose(perm = var_656_perm_0, x = var_655_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [130, 896]> input_121_cast_fp16 = reshape(shape = var_657, x = var_656_cast_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_9_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100555008))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101357888))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_9_out_bias_to_fp16 = const()[name = string("trunk_layers_9_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101359744)))];
            tensor<fp16, [130, 896]> linear_58_cast_fp16 = linear(bias = trunk_layers_9_out_bias_to_fp16, weight = trunk_layers_9_out_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<fp16, [130, 896]> input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_58_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<int32, [1]> input_125_axes_0 = const()[name = string("input_125_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_9_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_9_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101361600)))];
            tensor<fp16, [896]> trunk_layers_9_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_9_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101363456)))];
            tensor<fp16, [130, 896]> input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = trunk_layers_9_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_9_mlp_ln_weight_to_fp16, x = input_123_cast_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_9_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101365312))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104576640))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_9_fc1_bias_to_fp16 = const()[name = string("trunk_layers_9_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104583872)))];
            tensor<fp16, [130, 3584]> linear_59_cast_fp16 = linear(bias = trunk_layers_9_fc1_bias_to_fp16, weight = trunk_layers_9_fc1_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_59_cast_fp16")];
            string input_127_mode_0 = const()[name = string("input_127_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_127_cast_fp16 = gelu(mode = input_127_mode_0, x = linear_59_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_9_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104591104))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107802432))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_9_fc2_bias_to_fp16 = const()[name = string("trunk_layers_9_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107804288)))];
            tensor<fp16, [130, 896]> linear_60_cast_fp16 = linear(bias = trunk_layers_9_fc2_bias_to_fp16, weight = trunk_layers_9_fc2_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [130, 896]> input_129_cast_fp16 = add(x = input_123_cast_fp16, y = linear_60_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<int32, [1]> input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_10_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_10_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107806144)))];
            tensor<fp16, [896]> trunk_layers_10_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_10_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107808000)))];
            tensor<fp16, [130, 896]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = trunk_layers_10_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_10_attn_ln_weight_to_fp16, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_10_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107809856))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108612736))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_10_q_bias_to_fp16 = const()[name = string("trunk_layers_10_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108614592)))];
            tensor<fp16, [130, 896]> linear_61_cast_fp16 = linear(bias = trunk_layers_10_q_bias_to_fp16, weight = trunk_layers_10_q_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<int32, [3]> var_691 = const()[name = string("op_691"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_692_cast_fp16 = reshape(shape = var_691, x = linear_61_cast_fp16)[name = string("op_692_cast_fp16")];
            tensor<int32, [3]> var_693_perm_0 = const()[name = string("op_693_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_21_axes_0 = const()[name = string("q_21_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_693_cast_fp16 = transpose(perm = var_693_perm_0, x = var_692_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 14, 130, 64]> q_21_cast_fp16 = expand_dims(axes = q_21_axes_0, x = var_693_cast_fp16)[name = string("q_21_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_10_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108616448))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109419328))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_10_k_bias_to_fp16 = const()[name = string("trunk_layers_10_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109421184)))];
            tensor<fp16, [130, 896]> linear_62_cast_fp16 = linear(bias = trunk_layers_10_k_bias_to_fp16, weight = trunk_layers_10_k_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<int32, [3]> var_698 = const()[name = string("op_698"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_699_cast_fp16 = reshape(shape = var_698, x = linear_62_cast_fp16)[name = string("op_699_cast_fp16")];
            tensor<int32, [3]> var_700_perm_0 = const()[name = string("op_700_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_21_axes_0 = const()[name = string("k_21_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_700_cast_fp16 = transpose(perm = var_700_perm_0, x = var_699_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 14, 130, 64]> k_21_cast_fp16 = expand_dims(axes = k_21_axes_0, x = var_700_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_10_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109423040))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110225920))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_10_v_bias_to_fp16 = const()[name = string("trunk_layers_10_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110227776)))];
            tensor<fp16, [130, 896]> linear_63_cast_fp16 = linear(bias = trunk_layers_10_v_bias_to_fp16, weight = trunk_layers_10_v_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [3]> var_705 = const()[name = string("op_705"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_706_cast_fp16 = reshape(shape = var_705, x = linear_63_cast_fp16)[name = string("op_706_cast_fp16")];
            tensor<int32, [3]> var_707_perm_0 = const()[name = string("op_707_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_21_axes_0 = const()[name = string("v_21_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_707_cast_fp16 = transpose(perm = var_707_perm_0, x = var_706_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 14, 130, 64]> v_21_cast_fp16 = expand_dims(axes = v_21_axes_0, x = var_707_cast_fp16)[name = string("v_21_cast_fp16")];
            fp16 mul_10_y_0_to_fp16 = const()[name = string("mul_10_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_10_cast_fp16 = mul(x = q_21_cast_fp16, y = mul_10_y_0_to_fp16)[name = string("mul_10_cast_fp16")];
            bool matmul_10_transpose_y_0 = const()[name = string("matmul_10_transpose_y_0"), val = bool(true)];
            bool matmul_10_transpose_x_0 = const()[name = string("matmul_10_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_10_cast_fp16, y = k_21_cast_fp16)[name = string("matmul_10_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_10_cast_fp16 = add(x = matmul_10_cast_fp16, y = attn_mask_to_fp16)[name = string("add_10_cast_fp16")];
            int32 softmax_10_axis_0 = const()[name = string("softmax_10_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = add_10_cast_fp16)[name = string("softmax_10_cast_fp16")];
            bool a_21_transpose_x_0 = const()[name = string("a_21_transpose_x_0"), val = bool(false)];
            bool a_21_transpose_y_0 = const()[name = string("a_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_21_cast_fp16 = matmul(transpose_x = a_21_transpose_x_0, transpose_y = a_21_transpose_y_0, x = softmax_10_cast_fp16, y = v_21_cast_fp16)[name = string("a_21_cast_fp16")];
            tensor<int32, [1]> var_710_axes_0 = const()[name = string("op_710_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_710_cast_fp16 = squeeze(axes = var_710_axes_0, x = a_21_cast_fp16)[name = string("op_710_cast_fp16")];
            tensor<int32, [3]> var_711_perm_0 = const()[name = string("op_711_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_712 = const()[name = string("op_712"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_711_cast_fp16 = transpose(perm = var_711_perm_0, x = var_710_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [130, 896]> input_133_cast_fp16 = reshape(shape = var_712, x = var_711_cast_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_10_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110229632))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111032512))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_10_out_bias_to_fp16 = const()[name = string("trunk_layers_10_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111034368)))];
            tensor<fp16, [130, 896]> linear_64_cast_fp16 = linear(bias = trunk_layers_10_out_bias_to_fp16, weight = trunk_layers_10_out_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<fp16, [130, 896]> input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_64_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<int32, [1]> input_137_axes_0 = const()[name = string("input_137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_10_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_10_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111036224)))];
            tensor<fp16, [896]> trunk_layers_10_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_10_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111038080)))];
            tensor<fp16, [130, 896]> input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = trunk_layers_10_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_10_mlp_ln_weight_to_fp16, x = input_135_cast_fp16)[name = string("input_137_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_10_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111039936))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114251264))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_10_fc1_bias_to_fp16 = const()[name = string("trunk_layers_10_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114258496)))];
            tensor<fp16, [130, 3584]> linear_65_cast_fp16 = linear(bias = trunk_layers_10_fc1_bias_to_fp16, weight = trunk_layers_10_fc1_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_65_cast_fp16")];
            string input_139_mode_0 = const()[name = string("input_139_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_139_cast_fp16 = gelu(mode = input_139_mode_0, x = linear_65_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_10_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114265728))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117477056))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_10_fc2_bias_to_fp16 = const()[name = string("trunk_layers_10_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117478912)))];
            tensor<fp16, [130, 896]> linear_66_cast_fp16 = linear(bias = trunk_layers_10_fc2_bias_to_fp16, weight = trunk_layers_10_fc2_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [130, 896]> input_141_cast_fp16 = add(x = input_135_cast_fp16, y = linear_66_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<int32, [1]> input_143_axes_0 = const()[name = string("input_143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_11_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_11_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117480768)))];
            tensor<fp16, [896]> trunk_layers_11_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_11_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117482624)))];
            tensor<fp16, [130, 896]> input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = trunk_layers_11_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_11_attn_ln_weight_to_fp16, x = input_141_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_11_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117484480))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118287360))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_11_q_bias_to_fp16 = const()[name = string("trunk_layers_11_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118289216)))];
            tensor<fp16, [130, 896]> linear_67_cast_fp16 = linear(bias = trunk_layers_11_q_bias_to_fp16, weight = trunk_layers_11_q_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<int32, [3]> var_746 = const()[name = string("op_746"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_747_cast_fp16 = reshape(shape = var_746, x = linear_67_cast_fp16)[name = string("op_747_cast_fp16")];
            tensor<int32, [3]> var_748_perm_0 = const()[name = string("op_748_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_23_axes_0 = const()[name = string("q_23_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_748_cast_fp16 = transpose(perm = var_748_perm_0, x = var_747_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 14, 130, 64]> q_23_cast_fp16 = expand_dims(axes = q_23_axes_0, x = var_748_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_11_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118291072))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119093952))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_11_k_bias_to_fp16 = const()[name = string("trunk_layers_11_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119095808)))];
            tensor<fp16, [130, 896]> linear_68_cast_fp16 = linear(bias = trunk_layers_11_k_bias_to_fp16, weight = trunk_layers_11_k_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<int32, [3]> var_753 = const()[name = string("op_753"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_754_cast_fp16 = reshape(shape = var_753, x = linear_68_cast_fp16)[name = string("op_754_cast_fp16")];
            tensor<int32, [3]> var_755_perm_0 = const()[name = string("op_755_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_23_axes_0 = const()[name = string("k_23_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_755_cast_fp16 = transpose(perm = var_755_perm_0, x = var_754_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 14, 130, 64]> k_23_cast_fp16 = expand_dims(axes = k_23_axes_0, x = var_755_cast_fp16)[name = string("k_23_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_11_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119097664))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119900544))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_11_v_bias_to_fp16 = const()[name = string("trunk_layers_11_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119902400)))];
            tensor<fp16, [130, 896]> linear_69_cast_fp16 = linear(bias = trunk_layers_11_v_bias_to_fp16, weight = trunk_layers_11_v_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<int32, [3]> var_760 = const()[name = string("op_760"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_761_cast_fp16 = reshape(shape = var_760, x = linear_69_cast_fp16)[name = string("op_761_cast_fp16")];
            tensor<int32, [3]> var_762_perm_0 = const()[name = string("op_762_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_23_axes_0 = const()[name = string("v_23_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_762_cast_fp16 = transpose(perm = var_762_perm_0, x = var_761_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 14, 130, 64]> v_23_cast_fp16 = expand_dims(axes = v_23_axes_0, x = var_762_cast_fp16)[name = string("v_23_cast_fp16")];
            fp16 mul_11_y_0_to_fp16 = const()[name = string("mul_11_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_11_cast_fp16 = mul(x = q_23_cast_fp16, y = mul_11_y_0_to_fp16)[name = string("mul_11_cast_fp16")];
            bool matmul_11_transpose_y_0 = const()[name = string("matmul_11_transpose_y_0"), val = bool(true)];
            bool matmul_11_transpose_x_0 = const()[name = string("matmul_11_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_11_cast_fp16, y = k_23_cast_fp16)[name = string("matmul_11_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_11_cast_fp16 = add(x = matmul_11_cast_fp16, y = attn_mask_to_fp16)[name = string("add_11_cast_fp16")];
            int32 softmax_11_axis_0 = const()[name = string("softmax_11_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = add_11_cast_fp16)[name = string("softmax_11_cast_fp16")];
            bool a_23_transpose_x_0 = const()[name = string("a_23_transpose_x_0"), val = bool(false)];
            bool a_23_transpose_y_0 = const()[name = string("a_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_23_cast_fp16 = matmul(transpose_x = a_23_transpose_x_0, transpose_y = a_23_transpose_y_0, x = softmax_11_cast_fp16, y = v_23_cast_fp16)[name = string("a_23_cast_fp16")];
            tensor<int32, [1]> var_765_axes_0 = const()[name = string("op_765_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_765_cast_fp16 = squeeze(axes = var_765_axes_0, x = a_23_cast_fp16)[name = string("op_765_cast_fp16")];
            tensor<int32, [3]> var_766_perm_0 = const()[name = string("op_766_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_767 = const()[name = string("op_767"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_766_cast_fp16 = transpose(perm = var_766_perm_0, x = var_765_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [130, 896]> input_145_cast_fp16 = reshape(shape = var_767, x = var_766_cast_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_11_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119904256))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120707136))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_11_out_bias_to_fp16 = const()[name = string("trunk_layers_11_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120708992)))];
            tensor<fp16, [130, 896]> linear_70_cast_fp16 = linear(bias = trunk_layers_11_out_bias_to_fp16, weight = trunk_layers_11_out_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [130, 896]> input_147_cast_fp16 = add(x = input_141_cast_fp16, y = linear_70_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<int32, [1]> input_149_axes_0 = const()[name = string("input_149_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_11_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_11_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120710848)))];
            tensor<fp16, [896]> trunk_layers_11_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_11_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120712704)))];
            tensor<fp16, [130, 896]> input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = trunk_layers_11_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_11_mlp_ln_weight_to_fp16, x = input_147_cast_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_11_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120714560))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123925888))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_11_fc1_bias_to_fp16 = const()[name = string("trunk_layers_11_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123933120)))];
            tensor<fp16, [130, 3584]> linear_71_cast_fp16 = linear(bias = trunk_layers_11_fc1_bias_to_fp16, weight = trunk_layers_11_fc1_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_71_cast_fp16")];
            string input_151_mode_0 = const()[name = string("input_151_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = linear_71_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_11_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123940352))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127151680))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_11_fc2_bias_to_fp16 = const()[name = string("trunk_layers_11_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127153536)))];
            tensor<fp16, [130, 896]> linear_72_cast_fp16 = linear(bias = trunk_layers_11_fc2_bias_to_fp16, weight = trunk_layers_11_fc2_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<fp16, [130, 896]> input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_72_cast_fp16)[name = string("input_153_cast_fp16")];
            tensor<int32, [1]> input_155_axes_0 = const()[name = string("input_155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_12_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_12_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127155392)))];
            tensor<fp16, [896]> trunk_layers_12_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_12_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127157248)))];
            tensor<fp16, [130, 896]> input_155_cast_fp16 = layer_norm(axes = input_155_axes_0, beta = trunk_layers_12_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_12_attn_ln_weight_to_fp16, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_12_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127159104))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127961984))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_12_q_bias_to_fp16 = const()[name = string("trunk_layers_12_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127963840)))];
            tensor<fp16, [130, 896]> linear_73_cast_fp16 = linear(bias = trunk_layers_12_q_bias_to_fp16, weight = trunk_layers_12_q_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<int32, [3]> var_801 = const()[name = string("op_801"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_802_cast_fp16 = reshape(shape = var_801, x = linear_73_cast_fp16)[name = string("op_802_cast_fp16")];
            tensor<int32, [3]> var_803_perm_0 = const()[name = string("op_803_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_25_axes_0 = const()[name = string("q_25_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_803_cast_fp16 = transpose(perm = var_803_perm_0, x = var_802_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 14, 130, 64]> q_25_cast_fp16 = expand_dims(axes = q_25_axes_0, x = var_803_cast_fp16)[name = string("q_25_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_12_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127965696))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128768576))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_12_k_bias_to_fp16 = const()[name = string("trunk_layers_12_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128770432)))];
            tensor<fp16, [130, 896]> linear_74_cast_fp16 = linear(bias = trunk_layers_12_k_bias_to_fp16, weight = trunk_layers_12_k_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<int32, [3]> var_808 = const()[name = string("op_808"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_809_cast_fp16 = reshape(shape = var_808, x = linear_74_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<int32, [3]> var_810_perm_0 = const()[name = string("op_810_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_25_axes_0 = const()[name = string("k_25_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_810_cast_fp16 = transpose(perm = var_810_perm_0, x = var_809_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 14, 130, 64]> k_25_cast_fp16 = expand_dims(axes = k_25_axes_0, x = var_810_cast_fp16)[name = string("k_25_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_12_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128772288))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129575168))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_12_v_bias_to_fp16 = const()[name = string("trunk_layers_12_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129577024)))];
            tensor<fp16, [130, 896]> linear_75_cast_fp16 = linear(bias = trunk_layers_12_v_bias_to_fp16, weight = trunk_layers_12_v_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<int32, [3]> var_815 = const()[name = string("op_815"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_816_cast_fp16 = reshape(shape = var_815, x = linear_75_cast_fp16)[name = string("op_816_cast_fp16")];
            tensor<int32, [3]> var_817_perm_0 = const()[name = string("op_817_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_25_axes_0 = const()[name = string("v_25_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_817_cast_fp16 = transpose(perm = var_817_perm_0, x = var_816_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 14, 130, 64]> v_25_cast_fp16 = expand_dims(axes = v_25_axes_0, x = var_817_cast_fp16)[name = string("v_25_cast_fp16")];
            fp16 mul_12_y_0_to_fp16 = const()[name = string("mul_12_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_12_cast_fp16 = mul(x = q_25_cast_fp16, y = mul_12_y_0_to_fp16)[name = string("mul_12_cast_fp16")];
            bool matmul_12_transpose_y_0 = const()[name = string("matmul_12_transpose_y_0"), val = bool(true)];
            bool matmul_12_transpose_x_0 = const()[name = string("matmul_12_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = mul_12_cast_fp16, y = k_25_cast_fp16)[name = string("matmul_12_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_12_cast_fp16 = add(x = matmul_12_cast_fp16, y = attn_mask_to_fp16)[name = string("add_12_cast_fp16")];
            int32 softmax_12_axis_0 = const()[name = string("softmax_12_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = add_12_cast_fp16)[name = string("softmax_12_cast_fp16")];
            bool a_25_transpose_x_0 = const()[name = string("a_25_transpose_x_0"), val = bool(false)];
            bool a_25_transpose_y_0 = const()[name = string("a_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_25_cast_fp16 = matmul(transpose_x = a_25_transpose_x_0, transpose_y = a_25_transpose_y_0, x = softmax_12_cast_fp16, y = v_25_cast_fp16)[name = string("a_25_cast_fp16")];
            tensor<int32, [1]> var_820_axes_0 = const()[name = string("op_820_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_820_cast_fp16 = squeeze(axes = var_820_axes_0, x = a_25_cast_fp16)[name = string("op_820_cast_fp16")];
            tensor<int32, [3]> var_821_perm_0 = const()[name = string("op_821_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_822 = const()[name = string("op_822"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_821_cast_fp16 = transpose(perm = var_821_perm_0, x = var_820_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [130, 896]> input_157_cast_fp16 = reshape(shape = var_822, x = var_821_cast_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_12_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129578880))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130381760))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_12_out_bias_to_fp16 = const()[name = string("trunk_layers_12_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130383616)))];
            tensor<fp16, [130, 896]> linear_76_cast_fp16 = linear(bias = trunk_layers_12_out_bias_to_fp16, weight = trunk_layers_12_out_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [130, 896]> input_159_cast_fp16 = add(x = input_153_cast_fp16, y = linear_76_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<int32, [1]> input_161_axes_0 = const()[name = string("input_161_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_12_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_12_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130385472)))];
            tensor<fp16, [896]> trunk_layers_12_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_12_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130387328)))];
            tensor<fp16, [130, 896]> input_161_cast_fp16 = layer_norm(axes = input_161_axes_0, beta = trunk_layers_12_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_12_mlp_ln_weight_to_fp16, x = input_159_cast_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_12_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130389184))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133600512))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_12_fc1_bias_to_fp16 = const()[name = string("trunk_layers_12_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133607744)))];
            tensor<fp16, [130, 3584]> linear_77_cast_fp16 = linear(bias = trunk_layers_12_fc1_bias_to_fp16, weight = trunk_layers_12_fc1_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_77_cast_fp16")];
            string input_163_mode_0 = const()[name = string("input_163_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_163_cast_fp16 = gelu(mode = input_163_mode_0, x = linear_77_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_12_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133614976))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136826304))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_12_fc2_bias_to_fp16 = const()[name = string("trunk_layers_12_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136828160)))];
            tensor<fp16, [130, 896]> linear_78_cast_fp16 = linear(bias = trunk_layers_12_fc2_bias_to_fp16, weight = trunk_layers_12_fc2_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<fp16, [130, 896]> input_165_cast_fp16 = add(x = input_159_cast_fp16, y = linear_78_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<int32, [1]> input_167_axes_0 = const()[name = string("input_167_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_13_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_13_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136830016)))];
            tensor<fp16, [896]> trunk_layers_13_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_13_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136831872)))];
            tensor<fp16, [130, 896]> input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = trunk_layers_13_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_13_attn_ln_weight_to_fp16, x = input_165_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_13_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136833728))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137636608))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_13_q_bias_to_fp16 = const()[name = string("trunk_layers_13_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137638464)))];
            tensor<fp16, [130, 896]> linear_79_cast_fp16 = linear(bias = trunk_layers_13_q_bias_to_fp16, weight = trunk_layers_13_q_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [3]> var_856 = const()[name = string("op_856"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_857_cast_fp16 = reshape(shape = var_856, x = linear_79_cast_fp16)[name = string("op_857_cast_fp16")];
            tensor<int32, [3]> var_858_perm_0 = const()[name = string("op_858_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_27_axes_0 = const()[name = string("q_27_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_858_cast_fp16 = transpose(perm = var_858_perm_0, x = var_857_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 14, 130, 64]> q_27_cast_fp16 = expand_dims(axes = q_27_axes_0, x = var_858_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_13_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137640320))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138443200))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_13_k_bias_to_fp16 = const()[name = string("trunk_layers_13_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138445056)))];
            tensor<fp16, [130, 896]> linear_80_cast_fp16 = linear(bias = trunk_layers_13_k_bias_to_fp16, weight = trunk_layers_13_k_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<int32, [3]> var_863 = const()[name = string("op_863"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_864_cast_fp16 = reshape(shape = var_863, x = linear_80_cast_fp16)[name = string("op_864_cast_fp16")];
            tensor<int32, [3]> var_865_perm_0 = const()[name = string("op_865_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_27_axes_0 = const()[name = string("k_27_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_865_cast_fp16 = transpose(perm = var_865_perm_0, x = var_864_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 14, 130, 64]> k_27_cast_fp16 = expand_dims(axes = k_27_axes_0, x = var_865_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_13_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138446912))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139249792))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_13_v_bias_to_fp16 = const()[name = string("trunk_layers_13_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139251648)))];
            tensor<fp16, [130, 896]> linear_81_cast_fp16 = linear(bias = trunk_layers_13_v_bias_to_fp16, weight = trunk_layers_13_v_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<int32, [3]> var_870 = const()[name = string("op_870"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_871_cast_fp16 = reshape(shape = var_870, x = linear_81_cast_fp16)[name = string("op_871_cast_fp16")];
            tensor<int32, [3]> var_872_perm_0 = const()[name = string("op_872_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_27_axes_0 = const()[name = string("v_27_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_872_cast_fp16 = transpose(perm = var_872_perm_0, x = var_871_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 14, 130, 64]> v_27_cast_fp16 = expand_dims(axes = v_27_axes_0, x = var_872_cast_fp16)[name = string("v_27_cast_fp16")];
            fp16 mul_13_y_0_to_fp16 = const()[name = string("mul_13_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_13_cast_fp16 = mul(x = q_27_cast_fp16, y = mul_13_y_0_to_fp16)[name = string("mul_13_cast_fp16")];
            bool matmul_13_transpose_y_0 = const()[name = string("matmul_13_transpose_y_0"), val = bool(true)];
            bool matmul_13_transpose_x_0 = const()[name = string("matmul_13_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = mul_13_cast_fp16, y = k_27_cast_fp16)[name = string("matmul_13_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_13_cast_fp16 = add(x = matmul_13_cast_fp16, y = attn_mask_to_fp16)[name = string("add_13_cast_fp16")];
            int32 softmax_13_axis_0 = const()[name = string("softmax_13_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = add_13_cast_fp16)[name = string("softmax_13_cast_fp16")];
            bool a_27_transpose_x_0 = const()[name = string("a_27_transpose_x_0"), val = bool(false)];
            bool a_27_transpose_y_0 = const()[name = string("a_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_27_cast_fp16 = matmul(transpose_x = a_27_transpose_x_0, transpose_y = a_27_transpose_y_0, x = softmax_13_cast_fp16, y = v_27_cast_fp16)[name = string("a_27_cast_fp16")];
            tensor<int32, [1]> var_875_axes_0 = const()[name = string("op_875_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_875_cast_fp16 = squeeze(axes = var_875_axes_0, x = a_27_cast_fp16)[name = string("op_875_cast_fp16")];
            tensor<int32, [3]> var_876_perm_0 = const()[name = string("op_876_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_877 = const()[name = string("op_877"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_876_cast_fp16 = transpose(perm = var_876_perm_0, x = var_875_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [130, 896]> input_169_cast_fp16 = reshape(shape = var_877, x = var_876_cast_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_13_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139253504))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140056384))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_13_out_bias_to_fp16 = const()[name = string("trunk_layers_13_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140058240)))];
            tensor<fp16, [130, 896]> linear_82_cast_fp16 = linear(bias = trunk_layers_13_out_bias_to_fp16, weight = trunk_layers_13_out_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [130, 896]> input_171_cast_fp16 = add(x = input_165_cast_fp16, y = linear_82_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<int32, [1]> input_173_axes_0 = const()[name = string("input_173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_13_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_13_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140060096)))];
            tensor<fp16, [896]> trunk_layers_13_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_13_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140061952)))];
            tensor<fp16, [130, 896]> input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = trunk_layers_13_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_13_mlp_ln_weight_to_fp16, x = input_171_cast_fp16)[name = string("input_173_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_13_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140063808))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143275136))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_13_fc1_bias_to_fp16 = const()[name = string("trunk_layers_13_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143282368)))];
            tensor<fp16, [130, 3584]> linear_83_cast_fp16 = linear(bias = trunk_layers_13_fc1_bias_to_fp16, weight = trunk_layers_13_fc1_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_83_cast_fp16")];
            string input_175_mode_0 = const()[name = string("input_175_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_175_cast_fp16 = gelu(mode = input_175_mode_0, x = linear_83_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_13_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143289600))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146500928))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_13_fc2_bias_to_fp16 = const()[name = string("trunk_layers_13_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146502784)))];
            tensor<fp16, [130, 896]> linear_84_cast_fp16 = linear(bias = trunk_layers_13_fc2_bias_to_fp16, weight = trunk_layers_13_fc2_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<fp16, [130, 896]> input_177_cast_fp16 = add(x = input_171_cast_fp16, y = linear_84_cast_fp16)[name = string("input_177_cast_fp16")];
            tensor<int32, [1]> input_179_axes_0 = const()[name = string("input_179_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_14_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_14_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146504640)))];
            tensor<fp16, [896]> trunk_layers_14_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_14_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146506496)))];
            tensor<fp16, [130, 896]> input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = trunk_layers_14_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_14_attn_ln_weight_to_fp16, x = input_177_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_14_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146508352))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147311232))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_14_q_bias_to_fp16 = const()[name = string("trunk_layers_14_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147313088)))];
            tensor<fp16, [130, 896]> linear_85_cast_fp16 = linear(bias = trunk_layers_14_q_bias_to_fp16, weight = trunk_layers_14_q_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [3]> var_911 = const()[name = string("op_911"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_912_cast_fp16 = reshape(shape = var_911, x = linear_85_cast_fp16)[name = string("op_912_cast_fp16")];
            tensor<int32, [3]> var_913_perm_0 = const()[name = string("op_913_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_29_axes_0 = const()[name = string("q_29_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_913_cast_fp16 = transpose(perm = var_913_perm_0, x = var_912_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 14, 130, 64]> q_29_cast_fp16 = expand_dims(axes = q_29_axes_0, x = var_913_cast_fp16)[name = string("q_29_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_14_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147314944))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148117824))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_14_k_bias_to_fp16 = const()[name = string("trunk_layers_14_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148119680)))];
            tensor<fp16, [130, 896]> linear_86_cast_fp16 = linear(bias = trunk_layers_14_k_bias_to_fp16, weight = trunk_layers_14_k_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [3]> var_918 = const()[name = string("op_918"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_919_cast_fp16 = reshape(shape = var_918, x = linear_86_cast_fp16)[name = string("op_919_cast_fp16")];
            tensor<int32, [3]> var_920_perm_0 = const()[name = string("op_920_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_29_axes_0 = const()[name = string("k_29_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_920_cast_fp16 = transpose(perm = var_920_perm_0, x = var_919_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 14, 130, 64]> k_29_cast_fp16 = expand_dims(axes = k_29_axes_0, x = var_920_cast_fp16)[name = string("k_29_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_14_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148121536))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148924416))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_14_v_bias_to_fp16 = const()[name = string("trunk_layers_14_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148926272)))];
            tensor<fp16, [130, 896]> linear_87_cast_fp16 = linear(bias = trunk_layers_14_v_bias_to_fp16, weight = trunk_layers_14_v_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<int32, [3]> var_925 = const()[name = string("op_925"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_926_cast_fp16 = reshape(shape = var_925, x = linear_87_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [3]> var_927_perm_0 = const()[name = string("op_927_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_29_axes_0 = const()[name = string("v_29_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_927_cast_fp16 = transpose(perm = var_927_perm_0, x = var_926_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 14, 130, 64]> v_29_cast_fp16 = expand_dims(axes = v_29_axes_0, x = var_927_cast_fp16)[name = string("v_29_cast_fp16")];
            fp16 mul_14_y_0_to_fp16 = const()[name = string("mul_14_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_14_cast_fp16 = mul(x = q_29_cast_fp16, y = mul_14_y_0_to_fp16)[name = string("mul_14_cast_fp16")];
            bool matmul_14_transpose_y_0 = const()[name = string("matmul_14_transpose_y_0"), val = bool(true)];
            bool matmul_14_transpose_x_0 = const()[name = string("matmul_14_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = mul_14_cast_fp16, y = k_29_cast_fp16)[name = string("matmul_14_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_14_cast_fp16 = add(x = matmul_14_cast_fp16, y = attn_mask_to_fp16)[name = string("add_14_cast_fp16")];
            int32 softmax_14_axis_0 = const()[name = string("softmax_14_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = add_14_cast_fp16)[name = string("softmax_14_cast_fp16")];
            bool a_29_transpose_x_0 = const()[name = string("a_29_transpose_x_0"), val = bool(false)];
            bool a_29_transpose_y_0 = const()[name = string("a_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_29_cast_fp16 = matmul(transpose_x = a_29_transpose_x_0, transpose_y = a_29_transpose_y_0, x = softmax_14_cast_fp16, y = v_29_cast_fp16)[name = string("a_29_cast_fp16")];
            tensor<int32, [1]> var_930_axes_0 = const()[name = string("op_930_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_930_cast_fp16 = squeeze(axes = var_930_axes_0, x = a_29_cast_fp16)[name = string("op_930_cast_fp16")];
            tensor<int32, [3]> var_931_perm_0 = const()[name = string("op_931_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_932 = const()[name = string("op_932"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_931_cast_fp16 = transpose(perm = var_931_perm_0, x = var_930_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [130, 896]> input_181_cast_fp16 = reshape(shape = var_932, x = var_931_cast_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_14_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148928128))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149731008))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_14_out_bias_to_fp16 = const()[name = string("trunk_layers_14_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149732864)))];
            tensor<fp16, [130, 896]> linear_88_cast_fp16 = linear(bias = trunk_layers_14_out_bias_to_fp16, weight = trunk_layers_14_out_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [130, 896]> input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_88_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<int32, [1]> input_185_axes_0 = const()[name = string("input_185_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_14_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_14_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149734720)))];
            tensor<fp16, [896]> trunk_layers_14_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_14_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149736576)))];
            tensor<fp16, [130, 896]> input_185_cast_fp16 = layer_norm(axes = input_185_axes_0, beta = trunk_layers_14_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_14_mlp_ln_weight_to_fp16, x = input_183_cast_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_14_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149738432))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152949760))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_14_fc1_bias_to_fp16 = const()[name = string("trunk_layers_14_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152956992)))];
            tensor<fp16, [130, 3584]> linear_89_cast_fp16 = linear(bias = trunk_layers_14_fc1_bias_to_fp16, weight = trunk_layers_14_fc1_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_89_cast_fp16")];
            string input_187_mode_0 = const()[name = string("input_187_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_187_cast_fp16 = gelu(mode = input_187_mode_0, x = linear_89_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_14_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152964224))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156175552))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_14_fc2_bias_to_fp16 = const()[name = string("trunk_layers_14_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156177408)))];
            tensor<fp16, [130, 896]> linear_90_cast_fp16 = linear(bias = trunk_layers_14_fc2_bias_to_fp16, weight = trunk_layers_14_fc2_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [130, 896]> input_189_cast_fp16 = add(x = input_183_cast_fp16, y = linear_90_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<int32, [1]> input_191_axes_0 = const()[name = string("input_191_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_15_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_15_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156179264)))];
            tensor<fp16, [896]> trunk_layers_15_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_15_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156181120)))];
            tensor<fp16, [130, 896]> input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = trunk_layers_15_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_15_attn_ln_weight_to_fp16, x = input_189_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_15_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156182976))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156985856))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_15_q_bias_to_fp16 = const()[name = string("trunk_layers_15_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156987712)))];
            tensor<fp16, [130, 896]> linear_91_cast_fp16 = linear(bias = trunk_layers_15_q_bias_to_fp16, weight = trunk_layers_15_q_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [3]> var_966 = const()[name = string("op_966"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_967_cast_fp16 = reshape(shape = var_966, x = linear_91_cast_fp16)[name = string("op_967_cast_fp16")];
            tensor<int32, [3]> var_968_perm_0 = const()[name = string("op_968_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_31_axes_0 = const()[name = string("q_31_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_968_cast_fp16 = transpose(perm = var_968_perm_0, x = var_967_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 14, 130, 64]> q_31_cast_fp16 = expand_dims(axes = q_31_axes_0, x = var_968_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_15_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156989568))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157792448))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_15_k_bias_to_fp16 = const()[name = string("trunk_layers_15_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157794304)))];
            tensor<fp16, [130, 896]> linear_92_cast_fp16 = linear(bias = trunk_layers_15_k_bias_to_fp16, weight = trunk_layers_15_k_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [3]> var_973 = const()[name = string("op_973"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_974_cast_fp16 = reshape(shape = var_973, x = linear_92_cast_fp16)[name = string("op_974_cast_fp16")];
            tensor<int32, [3]> var_975_perm_0 = const()[name = string("op_975_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_31_axes_0 = const()[name = string("k_31_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_975_cast_fp16 = transpose(perm = var_975_perm_0, x = var_974_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 14, 130, 64]> k_31_cast_fp16 = expand_dims(axes = k_31_axes_0, x = var_975_cast_fp16)[name = string("k_31_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_15_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157796160))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158599040))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_15_v_bias_to_fp16 = const()[name = string("trunk_layers_15_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158600896)))];
            tensor<fp16, [130, 896]> linear_93_cast_fp16 = linear(bias = trunk_layers_15_v_bias_to_fp16, weight = trunk_layers_15_v_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [3]> var_980 = const()[name = string("op_980"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_981_cast_fp16 = reshape(shape = var_980, x = linear_93_cast_fp16)[name = string("op_981_cast_fp16")];
            tensor<int32, [3]> var_982_perm_0 = const()[name = string("op_982_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_31_axes_0 = const()[name = string("v_31_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_982_cast_fp16 = transpose(perm = var_982_perm_0, x = var_981_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 14, 130, 64]> v_31_cast_fp16 = expand_dims(axes = v_31_axes_0, x = var_982_cast_fp16)[name = string("v_31_cast_fp16")];
            fp16 mul_15_y_0_to_fp16 = const()[name = string("mul_15_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_15_cast_fp16 = mul(x = q_31_cast_fp16, y = mul_15_y_0_to_fp16)[name = string("mul_15_cast_fp16")];
            bool matmul_15_transpose_y_0 = const()[name = string("matmul_15_transpose_y_0"), val = bool(true)];
            bool matmul_15_transpose_x_0 = const()[name = string("matmul_15_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = mul_15_cast_fp16, y = k_31_cast_fp16)[name = string("matmul_15_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_15_cast_fp16 = add(x = matmul_15_cast_fp16, y = attn_mask_to_fp16)[name = string("add_15_cast_fp16")];
            int32 softmax_15_axis_0 = const()[name = string("softmax_15_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = add_15_cast_fp16)[name = string("softmax_15_cast_fp16")];
            bool a_31_transpose_x_0 = const()[name = string("a_31_transpose_x_0"), val = bool(false)];
            bool a_31_transpose_y_0 = const()[name = string("a_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_31_cast_fp16 = matmul(transpose_x = a_31_transpose_x_0, transpose_y = a_31_transpose_y_0, x = softmax_15_cast_fp16, y = v_31_cast_fp16)[name = string("a_31_cast_fp16")];
            tensor<int32, [1]> var_985_axes_0 = const()[name = string("op_985_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_985_cast_fp16 = squeeze(axes = var_985_axes_0, x = a_31_cast_fp16)[name = string("op_985_cast_fp16")];
            tensor<int32, [3]> var_986_perm_0 = const()[name = string("op_986_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_987 = const()[name = string("op_987"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_986_cast_fp16 = transpose(perm = var_986_perm_0, x = var_985_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [130, 896]> input_193_cast_fp16 = reshape(shape = var_987, x = var_986_cast_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_15_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158602752))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159405632))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_15_out_bias_to_fp16 = const()[name = string("trunk_layers_15_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159407488)))];
            tensor<fp16, [130, 896]> linear_94_cast_fp16 = linear(bias = trunk_layers_15_out_bias_to_fp16, weight = trunk_layers_15_out_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [130, 896]> input_195_cast_fp16 = add(x = input_189_cast_fp16, y = linear_94_cast_fp16)[name = string("input_195_cast_fp16")];
            tensor<int32, [1]> input_197_axes_0 = const()[name = string("input_197_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_15_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_15_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159409344)))];
            tensor<fp16, [896]> trunk_layers_15_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_15_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159411200)))];
            tensor<fp16, [130, 896]> input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = trunk_layers_15_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_15_mlp_ln_weight_to_fp16, x = input_195_cast_fp16)[name = string("input_197_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_15_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159413056))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162624384))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_15_fc1_bias_to_fp16 = const()[name = string("trunk_layers_15_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162631616)))];
            tensor<fp16, [130, 3584]> linear_95_cast_fp16 = linear(bias = trunk_layers_15_fc1_bias_to_fp16, weight = trunk_layers_15_fc1_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_95_cast_fp16")];
            string input_199_mode_0 = const()[name = string("input_199_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = linear_95_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_15_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162638848))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165850176))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_15_fc2_bias_to_fp16 = const()[name = string("trunk_layers_15_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165852032)))];
            tensor<fp16, [130, 896]> linear_96_cast_fp16 = linear(bias = trunk_layers_15_fc2_bias_to_fp16, weight = trunk_layers_15_fc2_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [130, 896]> input_201_cast_fp16 = add(x = input_195_cast_fp16, y = linear_96_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<int32, [1]> input_203_axes_0 = const()[name = string("input_203_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_16_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_16_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165853888)))];
            tensor<fp16, [896]> trunk_layers_16_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_16_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165855744)))];
            tensor<fp16, [130, 896]> input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = trunk_layers_16_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_16_attn_ln_weight_to_fp16, x = input_201_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_16_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165857600))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166660480))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_16_q_bias_to_fp16 = const()[name = string("trunk_layers_16_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166662336)))];
            tensor<fp16, [130, 896]> linear_97_cast_fp16 = linear(bias = trunk_layers_16_q_bias_to_fp16, weight = trunk_layers_16_q_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<int32, [3]> var_1021 = const()[name = string("op_1021"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_1022_cast_fp16 = reshape(shape = var_1021, x = linear_97_cast_fp16)[name = string("op_1022_cast_fp16")];
            tensor<int32, [3]> var_1023_perm_0 = const()[name = string("op_1023_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_33_axes_0 = const()[name = string("q_33_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_1023_cast_fp16 = transpose(perm = var_1023_perm_0, x = var_1022_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 14, 130, 64]> q_33_cast_fp16 = expand_dims(axes = q_33_axes_0, x = var_1023_cast_fp16)[name = string("q_33_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_16_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166664192))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167467072))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_16_k_bias_to_fp16 = const()[name = string("trunk_layers_16_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167468928)))];
            tensor<fp16, [130, 896]> linear_98_cast_fp16 = linear(bias = trunk_layers_16_k_bias_to_fp16, weight = trunk_layers_16_k_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [3]> var_1028 = const()[name = string("op_1028"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_1029_cast_fp16 = reshape(shape = var_1028, x = linear_98_cast_fp16)[name = string("op_1029_cast_fp16")];
            tensor<int32, [3]> var_1030_perm_0 = const()[name = string("op_1030_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_33_axes_0 = const()[name = string("k_33_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_1030_cast_fp16 = transpose(perm = var_1030_perm_0, x = var_1029_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 14, 130, 64]> k_33_cast_fp16 = expand_dims(axes = k_33_axes_0, x = var_1030_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_16_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167470784))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168273664))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_16_v_bias_to_fp16 = const()[name = string("trunk_layers_16_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168275520)))];
            tensor<fp16, [130, 896]> linear_99_cast_fp16 = linear(bias = trunk_layers_16_v_bias_to_fp16, weight = trunk_layers_16_v_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [3]> var_1035 = const()[name = string("op_1035"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_1036_cast_fp16 = reshape(shape = var_1035, x = linear_99_cast_fp16)[name = string("op_1036_cast_fp16")];
            tensor<int32, [3]> var_1037_perm_0 = const()[name = string("op_1037_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_33_axes_0 = const()[name = string("v_33_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_1037_cast_fp16 = transpose(perm = var_1037_perm_0, x = var_1036_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 14, 130, 64]> v_33_cast_fp16 = expand_dims(axes = v_33_axes_0, x = var_1037_cast_fp16)[name = string("v_33_cast_fp16")];
            fp16 mul_16_y_0_to_fp16 = const()[name = string("mul_16_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_16_cast_fp16 = mul(x = q_33_cast_fp16, y = mul_16_y_0_to_fp16)[name = string("mul_16_cast_fp16")];
            bool matmul_16_transpose_y_0 = const()[name = string("matmul_16_transpose_y_0"), val = bool(true)];
            bool matmul_16_transpose_x_0 = const()[name = string("matmul_16_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = mul_16_cast_fp16, y = k_33_cast_fp16)[name = string("matmul_16_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_16_cast_fp16 = add(x = matmul_16_cast_fp16, y = attn_mask_to_fp16)[name = string("add_16_cast_fp16")];
            int32 softmax_16_axis_0 = const()[name = string("softmax_16_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_16_cast_fp16 = softmax(axis = softmax_16_axis_0, x = add_16_cast_fp16)[name = string("softmax_16_cast_fp16")];
            bool a_33_transpose_x_0 = const()[name = string("a_33_transpose_x_0"), val = bool(false)];
            bool a_33_transpose_y_0 = const()[name = string("a_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_33_cast_fp16 = matmul(transpose_x = a_33_transpose_x_0, transpose_y = a_33_transpose_y_0, x = softmax_16_cast_fp16, y = v_33_cast_fp16)[name = string("a_33_cast_fp16")];
            tensor<int32, [1]> var_1040_axes_0 = const()[name = string("op_1040_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_1040_cast_fp16 = squeeze(axes = var_1040_axes_0, x = a_33_cast_fp16)[name = string("op_1040_cast_fp16")];
            tensor<int32, [3]> var_1041_perm_0 = const()[name = string("op_1041_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_1042 = const()[name = string("op_1042"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_1041_cast_fp16 = transpose(perm = var_1041_perm_0, x = var_1040_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [130, 896]> input_205_cast_fp16 = reshape(shape = var_1042, x = var_1041_cast_fp16)[name = string("input_205_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_16_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168277376))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169080256))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_16_out_bias_to_fp16 = const()[name = string("trunk_layers_16_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169082112)))];
            tensor<fp16, [130, 896]> linear_100_cast_fp16 = linear(bias = trunk_layers_16_out_bias_to_fp16, weight = trunk_layers_16_out_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<fp16, [130, 896]> input_207_cast_fp16 = add(x = input_201_cast_fp16, y = linear_100_cast_fp16)[name = string("input_207_cast_fp16")];
            tensor<int32, [1]> input_209_axes_0 = const()[name = string("input_209_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_16_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_16_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169083968)))];
            tensor<fp16, [896]> trunk_layers_16_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_16_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169085824)))];
            tensor<fp16, [130, 896]> input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = trunk_layers_16_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_16_mlp_ln_weight_to_fp16, x = input_207_cast_fp16)[name = string("input_209_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_16_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169087680))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172299008))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_16_fc1_bias_to_fp16 = const()[name = string("trunk_layers_16_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172306240)))];
            tensor<fp16, [130, 3584]> linear_101_cast_fp16 = linear(bias = trunk_layers_16_fc1_bias_to_fp16, weight = trunk_layers_16_fc1_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_101_cast_fp16")];
            string input_211_mode_0 = const()[name = string("input_211_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_211_cast_fp16 = gelu(mode = input_211_mode_0, x = linear_101_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_16_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172313472))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175524800))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_16_fc2_bias_to_fp16 = const()[name = string("trunk_layers_16_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175526656)))];
            tensor<fp16, [130, 896]> linear_102_cast_fp16 = linear(bias = trunk_layers_16_fc2_bias_to_fp16, weight = trunk_layers_16_fc2_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [130, 896]> input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_102_cast_fp16)[name = string("input_213_cast_fp16")];
            tensor<int32, [1]> input_215_axes_0 = const()[name = string("input_215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_17_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_17_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175528512)))];
            tensor<fp16, [896]> trunk_layers_17_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_17_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175530368)))];
            tensor<fp16, [130, 896]> input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = trunk_layers_17_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_17_attn_ln_weight_to_fp16, x = input_213_cast_fp16)[name = string("input_215_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_17_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175532224))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176335104))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_17_q_bias_to_fp16 = const()[name = string("trunk_layers_17_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176336960)))];
            tensor<fp16, [130, 896]> linear_103_cast_fp16 = linear(bias = trunk_layers_17_q_bias_to_fp16, weight = trunk_layers_17_q_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<int32, [3]> var_1076 = const()[name = string("op_1076"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_1077_cast_fp16 = reshape(shape = var_1076, x = linear_103_cast_fp16)[name = string("op_1077_cast_fp16")];
            tensor<int32, [3]> var_1078_perm_0 = const()[name = string("op_1078_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_35_axes_0 = const()[name = string("q_35_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_1078_cast_fp16 = transpose(perm = var_1078_perm_0, x = var_1077_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 14, 130, 64]> q_35_cast_fp16 = expand_dims(axes = q_35_axes_0, x = var_1078_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_17_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176338816))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177141696))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_17_k_bias_to_fp16 = const()[name = string("trunk_layers_17_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177143552)))];
            tensor<fp16, [130, 896]> linear_104_cast_fp16 = linear(bias = trunk_layers_17_k_bias_to_fp16, weight = trunk_layers_17_k_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<int32, [3]> var_1083 = const()[name = string("op_1083"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_1084_cast_fp16 = reshape(shape = var_1083, x = linear_104_cast_fp16)[name = string("op_1084_cast_fp16")];
            tensor<int32, [3]> var_1085_perm_0 = const()[name = string("op_1085_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_35_axes_0 = const()[name = string("k_35_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_1085_cast_fp16 = transpose(perm = var_1085_perm_0, x = var_1084_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 14, 130, 64]> k_35_cast_fp16 = expand_dims(axes = k_35_axes_0, x = var_1085_cast_fp16)[name = string("k_35_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_17_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177145408))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177948288))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_17_v_bias_to_fp16 = const()[name = string("trunk_layers_17_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177950144)))];
            tensor<fp16, [130, 896]> linear_105_cast_fp16 = linear(bias = trunk_layers_17_v_bias_to_fp16, weight = trunk_layers_17_v_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [3]> var_1090 = const()[name = string("op_1090"), val = tensor<int32, [3]>([130, 14, 64])];
            tensor<fp16, [130, 14, 64]> var_1091_cast_fp16 = reshape(shape = var_1090, x = linear_105_cast_fp16)[name = string("op_1091_cast_fp16")];
            tensor<int32, [3]> var_1092_perm_0 = const()[name = string("op_1092_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_35_axes_0 = const()[name = string("v_35_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_1092_cast_fp16 = transpose(perm = var_1092_perm_0, x = var_1091_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 14, 130, 64]> v_35_cast_fp16 = expand_dims(axes = v_35_axes_0, x = var_1092_cast_fp16)[name = string("v_35_cast_fp16")];
            fp16 mul_17_y_0_to_fp16 = const()[name = string("mul_17_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 130, 64]> mul_17_cast_fp16 = mul(x = q_35_cast_fp16, y = mul_17_y_0_to_fp16)[name = string("mul_17_cast_fp16")];
            bool matmul_17_transpose_y_0 = const()[name = string("matmul_17_transpose_y_0"), val = bool(true)];
            bool matmul_17_transpose_x_0 = const()[name = string("matmul_17_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 130]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = mul_17_cast_fp16, y = k_35_cast_fp16)[name = string("matmul_17_cast_fp16")];
            tensor<fp16, [1, 14, 130, 130]> add_17_cast_fp16 = add(x = matmul_17_cast_fp16, y = attn_mask_to_fp16)[name = string("add_17_cast_fp16")];
            int32 softmax_17_axis_0 = const()[name = string("softmax_17_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 130, 130]> softmax_17_cast_fp16 = softmax(axis = softmax_17_axis_0, x = add_17_cast_fp16)[name = string("softmax_17_cast_fp16")];
            bool a_transpose_x_0 = const()[name = string("a_transpose_x_0"), val = bool(false)];
            bool a_transpose_y_0 = const()[name = string("a_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 130, 64]> a_cast_fp16 = matmul(transpose_x = a_transpose_x_0, transpose_y = a_transpose_y_0, x = softmax_17_cast_fp16, y = v_35_cast_fp16)[name = string("a_cast_fp16")];
            tensor<int32, [1]> var_1095_axes_0 = const()[name = string("op_1095_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 130, 64]> var_1095_cast_fp16 = squeeze(axes = var_1095_axes_0, x = a_cast_fp16)[name = string("op_1095_cast_fp16")];
            tensor<int32, [3]> var_1096_perm_0 = const()[name = string("op_1096_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_1097 = const()[name = string("op_1097"), val = tensor<int32, [2]>([130, 896])];
            tensor<fp16, [130, 14, 64]> var_1096_cast_fp16 = transpose(perm = var_1096_perm_0, x = var_1095_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [130, 896]> input_217_cast_fp16 = reshape(shape = var_1097, x = var_1096_cast_fp16)[name = string("input_217_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_17_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177952000))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178754880))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_17_out_bias_to_fp16 = const()[name = string("trunk_layers_17_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178756736)))];
            tensor<fp16, [130, 896]> linear_106_cast_fp16 = linear(bias = trunk_layers_17_out_bias_to_fp16, weight = trunk_layers_17_out_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<fp16, [130, 896]> input_219_cast_fp16 = add(x = input_213_cast_fp16, y = linear_106_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<int32, [1]> input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_17_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_17_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178758592)))];
            tensor<fp16, [896]> trunk_layers_17_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_17_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178760448)))];
            tensor<fp16, [130, 896]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = trunk_layers_17_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_17_mlp_ln_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_17_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178762304))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181973632))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_17_fc1_bias_to_fp16 = const()[name = string("trunk_layers_17_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181980864)))];
            tensor<fp16, [130, 3584]> linear_107_cast_fp16 = linear(bias = trunk_layers_17_fc1_bias_to_fp16, weight = trunk_layers_17_fc1_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_107_cast_fp16")];
            string input_223_mode_0 = const()[name = string("input_223_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 3584]> input_223_cast_fp16 = gelu(mode = input_223_mode_0, x = linear_107_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_17_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181988096))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185199424))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_17_fc2_bias_to_fp16 = const()[name = string("trunk_layers_17_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185201280)))];
            tensor<fp16, [130, 896]> linear_108_cast_fp16 = linear(bias = trunk_layers_17_fc2_bias_to_fp16, weight = trunk_layers_17_fc2_weight_to_fp16_quantized, x = input_223_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [130, 896]> input_225_cast_fp16 = add(x = input_219_cast_fp16, y = linear_108_cast_fp16)[name = string("input_225_cast_fp16")];
            tensor<int32, [1]> input_227_axes_0 = const()[name = string("input_227_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_ln_post_weight_to_fp16 = const()[name = string("trunk_ln_post_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185203136)))];
            tensor<fp16, [896]> trunk_ln_post_bias_to_fp16 = const()[name = string("trunk_ln_post_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185204992)))];
            tensor<fp16, [130, 896]> input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = trunk_ln_post_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_ln_post_weight_to_fp16, x = input_225_cast_fp16)[name = string("input_227_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_proj1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_proj1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185206848))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186009728))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_proj1_bias_to_fp16 = const()[name = string("trunk_proj1_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186011584)))];
            tensor<fp16, [130, 896]> linear_109_cast_fp16 = linear(bias = trunk_proj1_bias_to_fp16, weight = trunk_proj1_weight_to_fp16_quantized, x = input_227_cast_fp16)[name = string("linear_109_cast_fp16")];
            string input_229_mode_0 = const()[name = string("input_229_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 896]> input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = linear_109_cast_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [1024, 896]> trunk_proj2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_proj2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186013440))), scale = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186932096))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186931008)))];
            tensor<fp16, [1024]> trunk_proj2_bias_to_fp16 = const()[name = string("trunk_proj2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186934208)))];
            tensor<fp16, [130, 1024]> linear_110_cast_fp16 = linear(bias = trunk_proj2_bias_to_fp16, weight = trunk_proj2_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<int32, [1]> input_233_axes_0 = const()[name = string("input_233_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> trunk_tower_0_norm_weight_to_fp16 = const()[name = string("trunk_tower_0_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186936320)))];
            tensor<fp16, [1024]> trunk_tower_0_norm_bias_to_fp16 = const()[name = string("trunk_tower_0_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186938432)))];
            tensor<fp16, [130, 1024]> input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = trunk_tower_0_norm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_tower_0_norm_weight_to_fp16, x = linear_110_cast_fp16)[name = string("input_233_cast_fp16")];
            tensor<fp16, [4096, 1024]> trunk_tower_0_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_0_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186940544))), scale = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191139072))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191134912)))];
            tensor<fp16, [4096]> trunk_tower_0_fc1_bias_to_fp16 = const()[name = string("trunk_tower_0_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191147328)))];
            tensor<fp16, [130, 4096]> linear_111_cast_fp16 = linear(bias = trunk_tower_0_fc1_bias_to_fp16, weight = trunk_tower_0_fc1_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_111_cast_fp16")];
            string input_235_mode_0 = const()[name = string("input_235_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 4096]> input_235_cast_fp16 = gelu(mode = input_235_mode_0, x = linear_111_cast_fp16)[name = string("input_235_cast_fp16")];
            tensor<fp16, [1024, 4096]> trunk_tower_0_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_0_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191155584))), scale = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195349952))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186931008)))];
            tensor<fp16, [1024]> trunk_tower_0_fc2_bias_to_fp16 = const()[name = string("trunk_tower_0_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195352064)))];
            tensor<fp16, [130, 1024]> linear_112_cast_fp16 = linear(bias = trunk_tower_0_fc2_bias_to_fp16, weight = trunk_tower_0_fc2_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<fp16, [130, 1024]> input_237_cast_fp16 = add(x = linear_110_cast_fp16, y = linear_112_cast_fp16)[name = string("input_237_cast_fp16")];
            tensor<int32, [1]> input_239_axes_0 = const()[name = string("input_239_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> trunk_tower_1_norm_weight_to_fp16 = const()[name = string("trunk_tower_1_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195354176)))];
            tensor<fp16, [1024]> trunk_tower_1_norm_bias_to_fp16 = const()[name = string("trunk_tower_1_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195356288)))];
            tensor<fp16, [130, 1024]> input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = trunk_tower_1_norm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_tower_1_norm_weight_to_fp16, x = input_237_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [4096, 1024]> trunk_tower_1_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_1_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195358400))), scale = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199552768))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191134912)))];
            tensor<fp16, [4096]> trunk_tower_1_fc1_bias_to_fp16 = const()[name = string("trunk_tower_1_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199561024)))];
            tensor<fp16, [130, 4096]> linear_113_cast_fp16 = linear(bias = trunk_tower_1_fc1_bias_to_fp16, weight = trunk_tower_1_fc1_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_113_cast_fp16")];
            string input_241_mode_0 = const()[name = string("input_241_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 4096]> input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_113_cast_fp16)[name = string("input_241_cast_fp16")];
            tensor<fp16, [1024, 4096]> trunk_tower_1_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_1_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199569280))), scale = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203763648))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186931008)))];
            tensor<fp16, [1024]> trunk_tower_1_fc2_bias_to_fp16 = const()[name = string("trunk_tower_1_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203765760)))];
            tensor<fp16, [130, 1024]> linear_114_cast_fp16 = linear(bias = trunk_tower_1_fc2_bias_to_fp16, weight = trunk_tower_1_fc2_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<fp16, [130, 1024]> input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_114_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<int32, [1]> input_245_axes_0 = const()[name = string("input_245_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> trunk_tower_2_norm_weight_to_fp16 = const()[name = string("trunk_tower_2_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203767872)))];
            tensor<fp16, [1024]> trunk_tower_2_norm_bias_to_fp16 = const()[name = string("trunk_tower_2_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203769984)))];
            tensor<fp16, [130, 1024]> input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = trunk_tower_2_norm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_tower_2_norm_weight_to_fp16, x = input_243_cast_fp16)[name = string("input_245_cast_fp16")];
            tensor<fp16, [4096, 1024]> trunk_tower_2_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_2_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203772096))), scale = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207966464))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191134912)))];
            tensor<fp16, [4096]> trunk_tower_2_fc1_bias_to_fp16 = const()[name = string("trunk_tower_2_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207974720)))];
            tensor<fp16, [130, 4096]> linear_115_cast_fp16 = linear(bias = trunk_tower_2_fc1_bias_to_fp16, weight = trunk_tower_2_fc1_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_115_cast_fp16")];
            string input_247_mode_0 = const()[name = string("input_247_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 4096]> input_247_cast_fp16 = gelu(mode = input_247_mode_0, x = linear_115_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<fp16, [1024, 4096]> trunk_tower_2_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_2_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207982976))), scale = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212177344))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186931008)))];
            tensor<fp16, [1024]> trunk_tower_2_fc2_bias_to_fp16 = const()[name = string("trunk_tower_2_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212179456)))];
            tensor<fp16, [130, 1024]> linear_116_cast_fp16 = linear(bias = trunk_tower_2_fc2_bias_to_fp16, weight = trunk_tower_2_fc2_weight_to_fp16_quantized, x = input_247_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [130, 1024]> input_249_cast_fp16 = add(x = input_243_cast_fp16, y = linear_116_cast_fp16)[name = string("input_249_cast_fp16")];
            tensor<int32, [1]> input_251_axes_0 = const()[name = string("input_251_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> trunk_tower_3_norm_weight_to_fp16 = const()[name = string("trunk_tower_3_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212181568)))];
            tensor<fp16, [1024]> trunk_tower_3_norm_bias_to_fp16 = const()[name = string("trunk_tower_3_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212183680)))];
            tensor<fp16, [130, 1024]> input_251_cast_fp16 = layer_norm(axes = input_251_axes_0, beta = trunk_tower_3_norm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_tower_3_norm_weight_to_fp16, x = input_249_cast_fp16)[name = string("input_251_cast_fp16")];
            tensor<fp16, [4096, 1024]> trunk_tower_3_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_3_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212185792))), scale = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216380160))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191134912)))];
            tensor<fp16, [4096]> trunk_tower_3_fc1_bias_to_fp16 = const()[name = string("trunk_tower_3_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216388416)))];
            tensor<fp16, [130, 4096]> linear_117_cast_fp16 = linear(bias = trunk_tower_3_fc1_bias_to_fp16, weight = trunk_tower_3_fc1_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_117_cast_fp16")];
            string input_253_mode_0 = const()[name = string("input_253_mode_0"), val = string("EXACT")];
            tensor<fp16, [130, 4096]> input_253_cast_fp16 = gelu(mode = input_253_mode_0, x = linear_117_cast_fp16)[name = string("input_253_cast_fp16")];
            tensor<fp16, [1024, 4096]> trunk_tower_3_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_3_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216396672))), scale = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220591040))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186931008)))];
            tensor<fp16, [1024]> trunk_tower_3_fc2_bias_to_fp16 = const()[name = string("trunk_tower_3_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220593152)))];
            tensor<fp16, [130, 1024]> linear_118_cast_fp16 = linear(bias = trunk_tower_3_fc2_bias_to_fp16, weight = trunk_tower_3_fc2_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [130, 1024]> input_cast_fp16 = add(x = input_249_cast_fp16, y = linear_118_cast_fp16)[name = string("input_cast_fp16")];
            tensor<int32, [1]> var_1189_axes_0 = const()[name = string("op_1189_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> trunk_final_norm_weight_to_fp16 = const()[name = string("trunk_final_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220595264)))];
            tensor<fp16, [1024]> trunk_final_norm_bias_to_fp16 = const()[name = string("trunk_final_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220597376)))];
            tensor<fp16, [130, 1024]> hidden = layer_norm(axes = var_1189_axes_0, beta = trunk_final_norm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_final_norm_weight_to_fp16, x = input_cast_fp16)[name = string("op_1189_cast_fp16")];
        } -> (hidden);
    func tower_30s<ios18>(tensor<fp32, [1, 1, 390, 390]> attn_mask, tensor<fp32, [1, 128, 3000]> audios) {
            tensor<int32, [3]> var_25_begin_0 = const()[name = string("op_25_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_25_end_0 = const()[name = string("op_25_end_0"), val = tensor<int32, [3]>([1, 128, 3000])];
            tensor<bool, [3]> var_25_end_mask_0 = const()[name = string("op_25_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> var_25_squeeze_mask_0 = const()[name = string("op_25_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            string audios_to_fp16_dtype_0 = const()[name = string("audios_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 3000]> audios_to_fp16 = cast(dtype = audios_to_fp16_dtype_0, x = audios)[name = string("cast_1")];
            tensor<fp16, [128, 3000]> var_25_cast_fp16 = slice_by_index(begin = var_25_begin_0, end = var_25_end_0, end_mask = var_25_end_mask_0, squeeze_mask = var_25_squeeze_mask_0, x = audios_to_fp16)[name = string("op_25_cast_fp16")];
            tensor<int32, [2]> x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor<int32, [2]>([1, 0])];
            tensor<int32, [3]> var_27 = const()[name = string("op_27"), val = tensor<int32, [3]>([30, 100, 128])];
            tensor<fp16, [3000, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = var_25_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [30, 100, 128]> x_3_cast_fp16 = reshape(shape = var_27, x = x_1_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<int32, [3]> var_29 = const()[name = string("op_29"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_1_axes_0 = const()[name = string("input_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [30, 128, 100]> var_30_cast_fp16 = transpose(perm = var_29, x = x_3_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [30, 1, 128, 100]> input_1_cast_fp16 = expand_dims(axes = input_1_axes_0, x = var_30_cast_fp16)[name = string("input_1_cast_fp16")];
            string var_38_pad_type_0 = const()[name = string("op_38_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_38_pad_0 = const()[name = string("op_38_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> var_38_strides_0 = const()[name = string("op_38_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_38_dilations_0 = const()[name = string("op_38_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_38_groups_0 = const()[name = string("op_38_groups_0"), val = int32(1)];
            tensor<fp16, [480, 1, 3, 3]> frontend_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("frontend_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [480, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5056))), zero_point = tensor<int8, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4480)))];
            tensor<fp16, [480]> frontend_conv1_bias_to_fp16 = const()[name = string("frontend_conv1_bias_to_fp16"), val = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6080)))];
            tensor<fp16, [30, 480, 64, 50]> var_38_cast_fp16 = conv(bias = frontend_conv1_bias_to_fp16, dilations = var_38_dilations_0, groups = var_38_groups_0, pad = var_38_pad_0, pad_type = var_38_pad_type_0, strides = var_38_strides_0, weight = frontend_conv1_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("op_38_cast_fp16")];
            string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("EXACT")];
            tensor<fp16, [30, 480, 64, 50]> input_3_cast_fp16 = gelu(mode = input_3_mode_0, x = var_38_cast_fp16)[name = string("input_3_cast_fp16")];
            string var_46_pad_type_0 = const()[name = string("op_46_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_46_pad_0 = const()[name = string("op_46_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> var_46_strides_0 = const()[name = string("op_46_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_46_dilations_0 = const()[name = string("op_46_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_46_groups_0 = const()[name = string("op_46_groups_0"), val = int32(1)];
            tensor<fp16, [480, 480, 3, 3]> frontend_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("frontend_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [480, 480, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7104))), scale = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2080768))), zero_point = tensor<int8, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4480)))];
            tensor<fp16, [480]> frontend_conv2_bias_to_fp16 = const()[name = string("frontend_conv2_bias_to_fp16"), val = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2081792)))];
            tensor<fp16, [30, 480, 32, 25]> var_46_cast_fp16 = conv(bias = frontend_conv2_bias_to_fp16, dilations = var_46_dilations_0, groups = var_46_groups_0, pad = var_46_pad_0, pad_type = var_46_pad_type_0, strides = var_46_strides_0, weight = frontend_conv2_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("op_46_cast_fp16")];
            string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("EXACT")];
            tensor<fp16, [30, 480, 32, 25]> input_5_cast_fp16 = gelu(mode = input_5_mode_0, x = var_46_cast_fp16)[name = string("input_5_cast_fp16")];
            string var_54_pad_type_0 = const()[name = string("op_54_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_54_pad_0 = const()[name = string("op_54_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> var_54_strides_0 = const()[name = string("op_54_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_54_dilations_0 = const()[name = string("op_54_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_54_groups_0 = const()[name = string("op_54_groups_0"), val = int32(1)];
            tensor<fp16, [480, 480, 3, 3]> frontend_conv3_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("frontend_conv3_weight_to_fp16_quantized"), quantized_data = tensor<int8, [480, 480, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2082816))), scale = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4156480))), zero_point = tensor<int8, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4480)))];
            tensor<fp16, [480]> frontend_conv3_bias_to_fp16 = const()[name = string("frontend_conv3_bias_to_fp16"), val = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4157504)))];
            tensor<fp16, [30, 480, 16, 13]> var_54_cast_fp16 = conv(bias = frontend_conv3_bias_to_fp16, dilations = var_54_dilations_0, groups = var_54_groups_0, pad = var_54_pad_0, pad_type = var_54_pad_type_0, strides = var_54_strides_0, weight = frontend_conv3_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("op_54_cast_fp16")];
            string x_5_mode_0 = const()[name = string("x_5_mode_0"), val = string("EXACT")];
            tensor<fp16, [30, 480, 16, 13]> x_5_cast_fp16 = gelu(mode = x_5_mode_0, x = var_54_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<int32, [4]> var_56 = const()[name = string("op_56"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [3]> var_58 = const()[name = string("op_58"), val = tensor<int32, [3]>([30, 13, 7680])];
            tensor<fp16, [30, 13, 480, 16]> var_57_cast_fp16 = transpose(perm = var_56, x = x_5_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [30, 13, 7680]> input_7_cast_fp16 = reshape(shape = var_58, x = var_57_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [896, 7680]> frontend_conv_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("frontend_conv_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 7680]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4158528))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11040832))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11042688)))];
            tensor<fp16, [30, 13, 896]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = frontend_conv_out_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<fp16, [1, 13, 896]> frontend_pos_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("frontend_pos_to_fp16_quantized"), quantized_data = tensor<int8, [1, 13, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11044544))), scale = fp16(0x1.02p-7), zero_point = int8(0)];
            tensor<fp16, [30, 13, 896]> x_cast_fp16 = add(x = linear_0_cast_fp16, y = frontend_pos_to_fp16_quantized)[name = string("x_cast_fp16")];
            tensor<int32, [2]> var_63 = const()[name = string("op_63"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 896]> input_9_cast_fp16 = reshape(shape = var_63, x = x_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<int32, [1]> input_11_axes_0 = const()[name = string("input_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_0_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_0_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11056256)))];
            tensor<fp16, [896]> trunk_layers_0_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_0_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11058112)))];
            fp16 var_67_to_fp16 = const()[name = string("op_67_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [390, 896]> input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = trunk_layers_0_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_0_attn_ln_weight_to_fp16, x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_0_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11059968))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11862848))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_0_q_bias_to_fp16 = const()[name = string("trunk_layers_0_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11864704)))];
            tensor<fp16, [390, 896]> linear_1_cast_fp16 = linear(bias = trunk_layers_0_q_bias_to_fp16, weight = trunk_layers_0_q_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [3]> var_141 = const()[name = string("op_141"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_142_cast_fp16 = reshape(shape = var_141, x = linear_1_cast_fp16)[name = string("op_142_cast_fp16")];
            tensor<int32, [3]> var_143_perm_0 = const()[name = string("op_143_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_1_axes_0 = const()[name = string("q_1_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_143_cast_fp16 = transpose(perm = var_143_perm_0, x = var_142_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 14, 390, 64]> q_1_cast_fp16 = expand_dims(axes = q_1_axes_0, x = var_143_cast_fp16)[name = string("q_1_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_0_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11866560))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12669440))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_0_k_bias_to_fp16 = const()[name = string("trunk_layers_0_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12671296)))];
            tensor<fp16, [390, 896]> linear_2_cast_fp16 = linear(bias = trunk_layers_0_k_bias_to_fp16, weight = trunk_layers_0_k_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [3]> var_148 = const()[name = string("op_148"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_149_cast_fp16 = reshape(shape = var_148, x = linear_2_cast_fp16)[name = string("op_149_cast_fp16")];
            tensor<int32, [3]> var_150_perm_0 = const()[name = string("op_150_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_1_axes_0 = const()[name = string("k_1_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_150_cast_fp16 = transpose(perm = var_150_perm_0, x = var_149_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 14, 390, 64]> k_1_cast_fp16 = expand_dims(axes = k_1_axes_0, x = var_150_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_0_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12673152))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13476032))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_0_v_bias_to_fp16 = const()[name = string("trunk_layers_0_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13477888)))];
            tensor<fp16, [390, 896]> linear_3_cast_fp16 = linear(bias = trunk_layers_0_v_bias_to_fp16, weight = trunk_layers_0_v_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<int32, [3]> var_155 = const()[name = string("op_155"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_156_cast_fp16 = reshape(shape = var_155, x = linear_3_cast_fp16)[name = string("op_156_cast_fp16")];
            tensor<int32, [3]> var_157_perm_0 = const()[name = string("op_157_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_1_axes_0 = const()[name = string("v_1_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_157_cast_fp16 = transpose(perm = var_157_perm_0, x = var_156_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 14, 390, 64]> v_1_cast_fp16 = expand_dims(axes = v_1_axes_0, x = var_157_cast_fp16)[name = string("v_1_cast_fp16")];
            fp16 mul_0_y_0_to_fp16 = const()[name = string("mul_0_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_0_cast_fp16 = mul(x = q_1_cast_fp16, y = mul_0_y_0_to_fp16)[name = string("mul_0_cast_fp16")];
            bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(true)];
            bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_0_cast_fp16, y = k_1_cast_fp16)[name = string("matmul_0_cast_fp16")];
            string attn_mask_to_fp16_dtype_0 = const()[name = string("attn_mask_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1, 390, 390]> attn_mask_to_fp16 = cast(dtype = attn_mask_to_fp16_dtype_0, x = attn_mask)[name = string("cast_0")];
            tensor<fp16, [1, 14, 390, 390]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = attn_mask_to_fp16)[name = string("add_0_cast_fp16")];
            int32 softmax_0_axis_0 = const()[name = string("softmax_0_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = string("softmax_0_cast_fp16")];
            bool a_1_transpose_x_0 = const()[name = string("a_1_transpose_x_0"), val = bool(false)];
            bool a_1_transpose_y_0 = const()[name = string("a_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_1_cast_fp16 = matmul(transpose_x = a_1_transpose_x_0, transpose_y = a_1_transpose_y_0, x = softmax_0_cast_fp16, y = v_1_cast_fp16)[name = string("a_1_cast_fp16")];
            tensor<int32, [1]> var_160_axes_0 = const()[name = string("op_160_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_160_cast_fp16 = squeeze(axes = var_160_axes_0, x = a_1_cast_fp16)[name = string("op_160_cast_fp16")];
            tensor<int32, [3]> var_161_perm_0 = const()[name = string("op_161_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_162 = const()[name = string("op_162"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_161_cast_fp16 = transpose(perm = var_161_perm_0, x = var_160_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [390, 896]> input_13_cast_fp16 = reshape(shape = var_162, x = var_161_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_0_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13479744))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14282624))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_0_out_bias_to_fp16 = const()[name = string("trunk_layers_0_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14284480)))];
            tensor<fp16, [390, 896]> linear_4_cast_fp16 = linear(bias = trunk_layers_0_out_bias_to_fp16, weight = trunk_layers_0_out_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [390, 896]> input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<int32, [1]> input_17_axes_0 = const()[name = string("input_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_0_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_0_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14286336)))];
            tensor<fp16, [896]> trunk_layers_0_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_0_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14288192)))];
            tensor<fp16, [390, 896]> input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = trunk_layers_0_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_0_mlp_ln_weight_to_fp16, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_0_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14290048))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17505024))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_0_fc1_bias_to_fp16 = const()[name = string("trunk_layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17512256)))];
            tensor<fp16, [390, 3584]> linear_5_cast_fp16 = linear(bias = trunk_layers_0_fc1_bias_to_fp16, weight = trunk_layers_0_fc1_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_5_cast_fp16")];
            string input_19_mode_0 = const()[name = string("input_19_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = linear_5_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_0_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17519488))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20730816))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_0_fc2_bias_to_fp16 = const()[name = string("trunk_layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20732672)))];
            tensor<fp16, [390, 896]> linear_6_cast_fp16 = linear(bias = trunk_layers_0_fc2_bias_to_fp16, weight = trunk_layers_0_fc2_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [390, 896]> input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<int32, [1]> input_23_axes_0 = const()[name = string("input_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_1_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_1_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20734528)))];
            tensor<fp16, [896]> trunk_layers_1_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_1_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20736384)))];
            tensor<fp16, [390, 896]> input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = trunk_layers_1_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_1_attn_ln_weight_to_fp16, x = input_21_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_1_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20738240))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21541120))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_1_q_bias_to_fp16 = const()[name = string("trunk_layers_1_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21542976)))];
            tensor<fp16, [390, 896]> linear_7_cast_fp16 = linear(bias = trunk_layers_1_q_bias_to_fp16, weight = trunk_layers_1_q_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [3]> var_196 = const()[name = string("op_196"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_197_cast_fp16 = reshape(shape = var_196, x = linear_7_cast_fp16)[name = string("op_197_cast_fp16")];
            tensor<int32, [3]> var_198_perm_0 = const()[name = string("op_198_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_3_axes_0 = const()[name = string("q_3_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_198_cast_fp16 = transpose(perm = var_198_perm_0, x = var_197_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 14, 390, 64]> q_3_cast_fp16 = expand_dims(axes = q_3_axes_0, x = var_198_cast_fp16)[name = string("q_3_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_1_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21544832))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22347712))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_1_k_bias_to_fp16 = const()[name = string("trunk_layers_1_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22349568)))];
            tensor<fp16, [390, 896]> linear_8_cast_fp16 = linear(bias = trunk_layers_1_k_bias_to_fp16, weight = trunk_layers_1_k_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [3]> var_203 = const()[name = string("op_203"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_204_cast_fp16 = reshape(shape = var_203, x = linear_8_cast_fp16)[name = string("op_204_cast_fp16")];
            tensor<int32, [3]> var_205_perm_0 = const()[name = string("op_205_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_3_axes_0 = const()[name = string("k_3_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_205_cast_fp16 = transpose(perm = var_205_perm_0, x = var_204_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 14, 390, 64]> k_3_cast_fp16 = expand_dims(axes = k_3_axes_0, x = var_205_cast_fp16)[name = string("k_3_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_1_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22351424))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23154304))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_1_v_bias_to_fp16 = const()[name = string("trunk_layers_1_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23156160)))];
            tensor<fp16, [390, 896]> linear_9_cast_fp16 = linear(bias = trunk_layers_1_v_bias_to_fp16, weight = trunk_layers_1_v_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [3]> var_210 = const()[name = string("op_210"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_211_cast_fp16 = reshape(shape = var_210, x = linear_9_cast_fp16)[name = string("op_211_cast_fp16")];
            tensor<int32, [3]> var_212_perm_0 = const()[name = string("op_212_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_3_axes_0 = const()[name = string("v_3_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_212_cast_fp16 = transpose(perm = var_212_perm_0, x = var_211_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 14, 390, 64]> v_3_cast_fp16 = expand_dims(axes = v_3_axes_0, x = var_212_cast_fp16)[name = string("v_3_cast_fp16")];
            fp16 mul_1_y_0_to_fp16 = const()[name = string("mul_1_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_1_cast_fp16 = mul(x = q_3_cast_fp16, y = mul_1_y_0_to_fp16)[name = string("mul_1_cast_fp16")];
            bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(true)];
            bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_1_cast_fp16, y = k_3_cast_fp16)[name = string("matmul_1_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = attn_mask_to_fp16)[name = string("add_1_cast_fp16")];
            int32 softmax_1_axis_0 = const()[name = string("softmax_1_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = string("softmax_1_cast_fp16")];
            bool a_3_transpose_x_0 = const()[name = string("a_3_transpose_x_0"), val = bool(false)];
            bool a_3_transpose_y_0 = const()[name = string("a_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_3_cast_fp16 = matmul(transpose_x = a_3_transpose_x_0, transpose_y = a_3_transpose_y_0, x = softmax_1_cast_fp16, y = v_3_cast_fp16)[name = string("a_3_cast_fp16")];
            tensor<int32, [1]> var_215_axes_0 = const()[name = string("op_215_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_215_cast_fp16 = squeeze(axes = var_215_axes_0, x = a_3_cast_fp16)[name = string("op_215_cast_fp16")];
            tensor<int32, [3]> var_216_perm_0 = const()[name = string("op_216_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_217 = const()[name = string("op_217"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_216_cast_fp16 = transpose(perm = var_216_perm_0, x = var_215_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [390, 896]> input_25_cast_fp16 = reshape(shape = var_217, x = var_216_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_1_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23158016))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23960896))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_1_out_bias_to_fp16 = const()[name = string("trunk_layers_1_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23962752)))];
            tensor<fp16, [390, 896]> linear_10_cast_fp16 = linear(bias = trunk_layers_1_out_bias_to_fp16, weight = trunk_layers_1_out_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [390, 896]> input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_10_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<int32, [1]> input_29_axes_0 = const()[name = string("input_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_1_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_1_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23964608)))];
            tensor<fp16, [896]> trunk_layers_1_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_1_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23966464)))];
            tensor<fp16, [390, 896]> input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = trunk_layers_1_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_1_mlp_ln_weight_to_fp16, x = input_27_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_1_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23968320))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27179648))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_1_fc1_bias_to_fp16 = const()[name = string("trunk_layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27186880)))];
            tensor<fp16, [390, 3584]> linear_11_cast_fp16 = linear(bias = trunk_layers_1_fc1_bias_to_fp16, weight = trunk_layers_1_fc1_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_11_cast_fp16")];
            string input_31_mode_0 = const()[name = string("input_31_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = linear_11_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_1_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27194112))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30405440))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_1_fc2_bias_to_fp16 = const()[name = string("trunk_layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30407296)))];
            tensor<fp16, [390, 896]> linear_12_cast_fp16 = linear(bias = trunk_layers_1_fc2_bias_to_fp16, weight = trunk_layers_1_fc2_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [390, 896]> input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_12_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<int32, [1]> input_35_axes_0 = const()[name = string("input_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_2_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_2_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30409152)))];
            tensor<fp16, [896]> trunk_layers_2_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_2_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30411008)))];
            tensor<fp16, [390, 896]> input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = trunk_layers_2_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_2_attn_ln_weight_to_fp16, x = input_33_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_2_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30412864))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31215744))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_2_q_bias_to_fp16 = const()[name = string("trunk_layers_2_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31217600)))];
            tensor<fp16, [390, 896]> linear_13_cast_fp16 = linear(bias = trunk_layers_2_q_bias_to_fp16, weight = trunk_layers_2_q_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<int32, [3]> var_251 = const()[name = string("op_251"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_252_cast_fp16 = reshape(shape = var_251, x = linear_13_cast_fp16)[name = string("op_252_cast_fp16")];
            tensor<int32, [3]> var_253_perm_0 = const()[name = string("op_253_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_5_axes_0 = const()[name = string("q_5_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_253_cast_fp16 = transpose(perm = var_253_perm_0, x = var_252_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 14, 390, 64]> q_5_cast_fp16 = expand_dims(axes = q_5_axes_0, x = var_253_cast_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_2_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31219456))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32022336))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_2_k_bias_to_fp16 = const()[name = string("trunk_layers_2_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32024192)))];
            tensor<fp16, [390, 896]> linear_14_cast_fp16 = linear(bias = trunk_layers_2_k_bias_to_fp16, weight = trunk_layers_2_k_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [3]> var_258 = const()[name = string("op_258"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_259_cast_fp16 = reshape(shape = var_258, x = linear_14_cast_fp16)[name = string("op_259_cast_fp16")];
            tensor<int32, [3]> var_260_perm_0 = const()[name = string("op_260_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_5_axes_0 = const()[name = string("k_5_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_260_cast_fp16 = transpose(perm = var_260_perm_0, x = var_259_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 14, 390, 64]> k_5_cast_fp16 = expand_dims(axes = k_5_axes_0, x = var_260_cast_fp16)[name = string("k_5_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_2_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32026048))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32828928))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_2_v_bias_to_fp16 = const()[name = string("trunk_layers_2_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32830784)))];
            tensor<fp16, [390, 896]> linear_15_cast_fp16 = linear(bias = trunk_layers_2_v_bias_to_fp16, weight = trunk_layers_2_v_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [3]> var_265 = const()[name = string("op_265"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_266_cast_fp16 = reshape(shape = var_265, x = linear_15_cast_fp16)[name = string("op_266_cast_fp16")];
            tensor<int32, [3]> var_267_perm_0 = const()[name = string("op_267_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_5_axes_0 = const()[name = string("v_5_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_267_cast_fp16 = transpose(perm = var_267_perm_0, x = var_266_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 14, 390, 64]> v_5_cast_fp16 = expand_dims(axes = v_5_axes_0, x = var_267_cast_fp16)[name = string("v_5_cast_fp16")];
            fp16 mul_2_y_0_to_fp16 = const()[name = string("mul_2_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_2_cast_fp16 = mul(x = q_5_cast_fp16, y = mul_2_y_0_to_fp16)[name = string("mul_2_cast_fp16")];
            bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(true)];
            bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_2_cast_fp16, y = k_5_cast_fp16)[name = string("matmul_2_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = attn_mask_to_fp16)[name = string("add_2_cast_fp16")];
            int32 softmax_2_axis_0 = const()[name = string("softmax_2_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16)[name = string("softmax_2_cast_fp16")];
            bool a_5_transpose_x_0 = const()[name = string("a_5_transpose_x_0"), val = bool(false)];
            bool a_5_transpose_y_0 = const()[name = string("a_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_5_cast_fp16 = matmul(transpose_x = a_5_transpose_x_0, transpose_y = a_5_transpose_y_0, x = softmax_2_cast_fp16, y = v_5_cast_fp16)[name = string("a_5_cast_fp16")];
            tensor<int32, [1]> var_270_axes_0 = const()[name = string("op_270_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_270_cast_fp16 = squeeze(axes = var_270_axes_0, x = a_5_cast_fp16)[name = string("op_270_cast_fp16")];
            tensor<int32, [3]> var_271_perm_0 = const()[name = string("op_271_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_272 = const()[name = string("op_272"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_271_cast_fp16 = transpose(perm = var_271_perm_0, x = var_270_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [390, 896]> input_37_cast_fp16 = reshape(shape = var_272, x = var_271_cast_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_2_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32832640))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33635520))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_2_out_bias_to_fp16 = const()[name = string("trunk_layers_2_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33637376)))];
            tensor<fp16, [390, 896]> linear_16_cast_fp16 = linear(bias = trunk_layers_2_out_bias_to_fp16, weight = trunk_layers_2_out_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<fp16, [390, 896]> input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_16_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<int32, [1]> input_41_axes_0 = const()[name = string("input_41_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_2_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_2_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33639232)))];
            tensor<fp16, [896]> trunk_layers_2_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_2_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33641088)))];
            tensor<fp16, [390, 896]> input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = trunk_layers_2_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_2_mlp_ln_weight_to_fp16, x = input_39_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_2_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33642944))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36854272))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_2_fc1_bias_to_fp16 = const()[name = string("trunk_layers_2_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36861504)))];
            tensor<fp16, [390, 3584]> linear_17_cast_fp16 = linear(bias = trunk_layers_2_fc1_bias_to_fp16, weight = trunk_layers_2_fc1_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_17_cast_fp16")];
            string input_43_mode_0 = const()[name = string("input_43_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_43_cast_fp16 = gelu(mode = input_43_mode_0, x = linear_17_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_2_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36868736))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40080064))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_2_fc2_bias_to_fp16 = const()[name = string("trunk_layers_2_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40081920)))];
            tensor<fp16, [390, 896]> linear_18_cast_fp16 = linear(bias = trunk_layers_2_fc2_bias_to_fp16, weight = trunk_layers_2_fc2_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [390, 896]> input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_18_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<int32, [1]> input_47_axes_0 = const()[name = string("input_47_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_3_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_3_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40083776)))];
            tensor<fp16, [896]> trunk_layers_3_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_3_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40085632)))];
            tensor<fp16, [390, 896]> input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = trunk_layers_3_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_3_attn_ln_weight_to_fp16, x = input_45_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_3_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40087488))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40890368))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_3_q_bias_to_fp16 = const()[name = string("trunk_layers_3_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40892224)))];
            tensor<fp16, [390, 896]> linear_19_cast_fp16 = linear(bias = trunk_layers_3_q_bias_to_fp16, weight = trunk_layers_3_q_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<int32, [3]> var_306 = const()[name = string("op_306"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_307_cast_fp16 = reshape(shape = var_306, x = linear_19_cast_fp16)[name = string("op_307_cast_fp16")];
            tensor<int32, [3]> var_308_perm_0 = const()[name = string("op_308_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_7_axes_0 = const()[name = string("q_7_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_308_cast_fp16 = transpose(perm = var_308_perm_0, x = var_307_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 14, 390, 64]> q_7_cast_fp16 = expand_dims(axes = q_7_axes_0, x = var_308_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_3_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40894080))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41696960))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_3_k_bias_to_fp16 = const()[name = string("trunk_layers_3_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41698816)))];
            tensor<fp16, [390, 896]> linear_20_cast_fp16 = linear(bias = trunk_layers_3_k_bias_to_fp16, weight = trunk_layers_3_k_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [3]> var_313 = const()[name = string("op_313"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_314_cast_fp16 = reshape(shape = var_313, x = linear_20_cast_fp16)[name = string("op_314_cast_fp16")];
            tensor<int32, [3]> var_315_perm_0 = const()[name = string("op_315_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_7_axes_0 = const()[name = string("k_7_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_315_cast_fp16 = transpose(perm = var_315_perm_0, x = var_314_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 14, 390, 64]> k_7_cast_fp16 = expand_dims(axes = k_7_axes_0, x = var_315_cast_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_3_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41700672))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42503552))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_3_v_bias_to_fp16 = const()[name = string("trunk_layers_3_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42505408)))];
            tensor<fp16, [390, 896]> linear_21_cast_fp16 = linear(bias = trunk_layers_3_v_bias_to_fp16, weight = trunk_layers_3_v_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [3]> var_320 = const()[name = string("op_320"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_321_cast_fp16 = reshape(shape = var_320, x = linear_21_cast_fp16)[name = string("op_321_cast_fp16")];
            tensor<int32, [3]> var_322_perm_0 = const()[name = string("op_322_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_7_axes_0 = const()[name = string("v_7_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_322_cast_fp16 = transpose(perm = var_322_perm_0, x = var_321_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 14, 390, 64]> v_7_cast_fp16 = expand_dims(axes = v_7_axes_0, x = var_322_cast_fp16)[name = string("v_7_cast_fp16")];
            fp16 mul_3_y_0_to_fp16 = const()[name = string("mul_3_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_3_cast_fp16 = mul(x = q_7_cast_fp16, y = mul_3_y_0_to_fp16)[name = string("mul_3_cast_fp16")];
            bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(true)];
            bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_3_cast_fp16, y = k_7_cast_fp16)[name = string("matmul_3_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = attn_mask_to_fp16)[name = string("add_3_cast_fp16")];
            int32 softmax_3_axis_0 = const()[name = string("softmax_3_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = add_3_cast_fp16)[name = string("softmax_3_cast_fp16")];
            bool a_7_transpose_x_0 = const()[name = string("a_7_transpose_x_0"), val = bool(false)];
            bool a_7_transpose_y_0 = const()[name = string("a_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_7_cast_fp16 = matmul(transpose_x = a_7_transpose_x_0, transpose_y = a_7_transpose_y_0, x = softmax_3_cast_fp16, y = v_7_cast_fp16)[name = string("a_7_cast_fp16")];
            tensor<int32, [1]> var_325_axes_0 = const()[name = string("op_325_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_325_cast_fp16 = squeeze(axes = var_325_axes_0, x = a_7_cast_fp16)[name = string("op_325_cast_fp16")];
            tensor<int32, [3]> var_326_perm_0 = const()[name = string("op_326_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_327 = const()[name = string("op_327"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_326_cast_fp16 = transpose(perm = var_326_perm_0, x = var_325_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [390, 896]> input_49_cast_fp16 = reshape(shape = var_327, x = var_326_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_3_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42507264))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43310144))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_3_out_bias_to_fp16 = const()[name = string("trunk_layers_3_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43312000)))];
            tensor<fp16, [390, 896]> linear_22_cast_fp16 = linear(bias = trunk_layers_3_out_bias_to_fp16, weight = trunk_layers_3_out_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<fp16, [390, 896]> input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_22_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<int32, [1]> input_53_axes_0 = const()[name = string("input_53_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_3_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_3_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43313856)))];
            tensor<fp16, [896]> trunk_layers_3_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_3_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43315712)))];
            tensor<fp16, [390, 896]> input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = trunk_layers_3_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_3_mlp_ln_weight_to_fp16, x = input_51_cast_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_3_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43317568))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46528896))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_3_fc1_bias_to_fp16 = const()[name = string("trunk_layers_3_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46536128)))];
            tensor<fp16, [390, 3584]> linear_23_cast_fp16 = linear(bias = trunk_layers_3_fc1_bias_to_fp16, weight = trunk_layers_3_fc1_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_23_cast_fp16")];
            string input_55_mode_0 = const()[name = string("input_55_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_55_cast_fp16 = gelu(mode = input_55_mode_0, x = linear_23_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_3_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46543360))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49754688))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_3_fc2_bias_to_fp16 = const()[name = string("trunk_layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49756544)))];
            tensor<fp16, [390, 896]> linear_24_cast_fp16 = linear(bias = trunk_layers_3_fc2_bias_to_fp16, weight = trunk_layers_3_fc2_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [390, 896]> input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_24_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<int32, [1]> input_59_axes_0 = const()[name = string("input_59_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_4_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_4_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49758400)))];
            tensor<fp16, [896]> trunk_layers_4_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_4_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49760256)))];
            tensor<fp16, [390, 896]> input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = trunk_layers_4_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_4_attn_ln_weight_to_fp16, x = input_57_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_4_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49762112))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50564992))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_4_q_bias_to_fp16 = const()[name = string("trunk_layers_4_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50566848)))];
            tensor<fp16, [390, 896]> linear_25_cast_fp16 = linear(bias = trunk_layers_4_q_bias_to_fp16, weight = trunk_layers_4_q_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<int32, [3]> var_361 = const()[name = string("op_361"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_362_cast_fp16 = reshape(shape = var_361, x = linear_25_cast_fp16)[name = string("op_362_cast_fp16")];
            tensor<int32, [3]> var_363_perm_0 = const()[name = string("op_363_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_9_axes_0 = const()[name = string("q_9_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_363_cast_fp16 = transpose(perm = var_363_perm_0, x = var_362_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 14, 390, 64]> q_9_cast_fp16 = expand_dims(axes = q_9_axes_0, x = var_363_cast_fp16)[name = string("q_9_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_4_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50568704))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51371584))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_4_k_bias_to_fp16 = const()[name = string("trunk_layers_4_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51373440)))];
            tensor<fp16, [390, 896]> linear_26_cast_fp16 = linear(bias = trunk_layers_4_k_bias_to_fp16, weight = trunk_layers_4_k_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<int32, [3]> var_368 = const()[name = string("op_368"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_369_cast_fp16 = reshape(shape = var_368, x = linear_26_cast_fp16)[name = string("op_369_cast_fp16")];
            tensor<int32, [3]> var_370_perm_0 = const()[name = string("op_370_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_9_axes_0 = const()[name = string("k_9_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_370_cast_fp16 = transpose(perm = var_370_perm_0, x = var_369_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 14, 390, 64]> k_9_cast_fp16 = expand_dims(axes = k_9_axes_0, x = var_370_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_4_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51375296))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52178176))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_4_v_bias_to_fp16 = const()[name = string("trunk_layers_4_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52180032)))];
            tensor<fp16, [390, 896]> linear_27_cast_fp16 = linear(bias = trunk_layers_4_v_bias_to_fp16, weight = trunk_layers_4_v_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<int32, [3]> var_375 = const()[name = string("op_375"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_376_cast_fp16 = reshape(shape = var_375, x = linear_27_cast_fp16)[name = string("op_376_cast_fp16")];
            tensor<int32, [3]> var_377_perm_0 = const()[name = string("op_377_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_9_axes_0 = const()[name = string("v_9_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_377_cast_fp16 = transpose(perm = var_377_perm_0, x = var_376_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 14, 390, 64]> v_9_cast_fp16 = expand_dims(axes = v_9_axes_0, x = var_377_cast_fp16)[name = string("v_9_cast_fp16")];
            fp16 mul_4_y_0_to_fp16 = const()[name = string("mul_4_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_4_cast_fp16 = mul(x = q_9_cast_fp16, y = mul_4_y_0_to_fp16)[name = string("mul_4_cast_fp16")];
            bool matmul_4_transpose_y_0 = const()[name = string("matmul_4_transpose_y_0"), val = bool(true)];
            bool matmul_4_transpose_x_0 = const()[name = string("matmul_4_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_4_cast_fp16, y = k_9_cast_fp16)[name = string("matmul_4_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = attn_mask_to_fp16)[name = string("add_4_cast_fp16")];
            int32 softmax_4_axis_0 = const()[name = string("softmax_4_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = add_4_cast_fp16)[name = string("softmax_4_cast_fp16")];
            bool a_9_transpose_x_0 = const()[name = string("a_9_transpose_x_0"), val = bool(false)];
            bool a_9_transpose_y_0 = const()[name = string("a_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_9_cast_fp16 = matmul(transpose_x = a_9_transpose_x_0, transpose_y = a_9_transpose_y_0, x = softmax_4_cast_fp16, y = v_9_cast_fp16)[name = string("a_9_cast_fp16")];
            tensor<int32, [1]> var_380_axes_0 = const()[name = string("op_380_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_380_cast_fp16 = squeeze(axes = var_380_axes_0, x = a_9_cast_fp16)[name = string("op_380_cast_fp16")];
            tensor<int32, [3]> var_381_perm_0 = const()[name = string("op_381_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_382 = const()[name = string("op_382"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_381_cast_fp16 = transpose(perm = var_381_perm_0, x = var_380_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [390, 896]> input_61_cast_fp16 = reshape(shape = var_382, x = var_381_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_4_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52181888))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52984768))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_4_out_bias_to_fp16 = const()[name = string("trunk_layers_4_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52986624)))];
            tensor<fp16, [390, 896]> linear_28_cast_fp16 = linear(bias = trunk_layers_4_out_bias_to_fp16, weight = trunk_layers_4_out_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<fp16, [390, 896]> input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_28_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<int32, [1]> input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_4_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_4_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52988480)))];
            tensor<fp16, [896]> trunk_layers_4_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_4_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52990336)))];
            tensor<fp16, [390, 896]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = trunk_layers_4_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_4_mlp_ln_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_4_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52992192))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56203520))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_4_fc1_bias_to_fp16 = const()[name = string("trunk_layers_4_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56210752)))];
            tensor<fp16, [390, 3584]> linear_29_cast_fp16 = linear(bias = trunk_layers_4_fc1_bias_to_fp16, weight = trunk_layers_4_fc1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_29_cast_fp16")];
            string input_67_mode_0 = const()[name = string("input_67_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_67_cast_fp16 = gelu(mode = input_67_mode_0, x = linear_29_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_4_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56217984))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59429312))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_4_fc2_bias_to_fp16 = const()[name = string("trunk_layers_4_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59431168)))];
            tensor<fp16, [390, 896]> linear_30_cast_fp16 = linear(bias = trunk_layers_4_fc2_bias_to_fp16, weight = trunk_layers_4_fc2_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<fp16, [390, 896]> input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_30_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<int32, [1]> input_71_axes_0 = const()[name = string("input_71_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_5_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_5_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59433024)))];
            tensor<fp16, [896]> trunk_layers_5_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_5_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59434880)))];
            tensor<fp16, [390, 896]> input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = trunk_layers_5_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_5_attn_ln_weight_to_fp16, x = input_69_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_5_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59436736))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60239616))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_5_q_bias_to_fp16 = const()[name = string("trunk_layers_5_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60241472)))];
            tensor<fp16, [390, 896]> linear_31_cast_fp16 = linear(bias = trunk_layers_5_q_bias_to_fp16, weight = trunk_layers_5_q_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<int32, [3]> var_416 = const()[name = string("op_416"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_417_cast_fp16 = reshape(shape = var_416, x = linear_31_cast_fp16)[name = string("op_417_cast_fp16")];
            tensor<int32, [3]> var_418_perm_0 = const()[name = string("op_418_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_11_axes_0 = const()[name = string("q_11_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_418_cast_fp16 = transpose(perm = var_418_perm_0, x = var_417_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 14, 390, 64]> q_11_cast_fp16 = expand_dims(axes = q_11_axes_0, x = var_418_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_5_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60243328))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61046208))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_5_k_bias_to_fp16 = const()[name = string("trunk_layers_5_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61048064)))];
            tensor<fp16, [390, 896]> linear_32_cast_fp16 = linear(bias = trunk_layers_5_k_bias_to_fp16, weight = trunk_layers_5_k_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<int32, [3]> var_423 = const()[name = string("op_423"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_424_cast_fp16 = reshape(shape = var_423, x = linear_32_cast_fp16)[name = string("op_424_cast_fp16")];
            tensor<int32, [3]> var_425_perm_0 = const()[name = string("op_425_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_11_axes_0 = const()[name = string("k_11_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_425_cast_fp16 = transpose(perm = var_425_perm_0, x = var_424_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 14, 390, 64]> k_11_cast_fp16 = expand_dims(axes = k_11_axes_0, x = var_425_cast_fp16)[name = string("k_11_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_5_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61049920))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61852800))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_5_v_bias_to_fp16 = const()[name = string("trunk_layers_5_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61854656)))];
            tensor<fp16, [390, 896]> linear_33_cast_fp16 = linear(bias = trunk_layers_5_v_bias_to_fp16, weight = trunk_layers_5_v_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<int32, [3]> var_430 = const()[name = string("op_430"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_431_cast_fp16 = reshape(shape = var_430, x = linear_33_cast_fp16)[name = string("op_431_cast_fp16")];
            tensor<int32, [3]> var_432_perm_0 = const()[name = string("op_432_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_11_axes_0 = const()[name = string("v_11_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_432_cast_fp16 = transpose(perm = var_432_perm_0, x = var_431_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 14, 390, 64]> v_11_cast_fp16 = expand_dims(axes = v_11_axes_0, x = var_432_cast_fp16)[name = string("v_11_cast_fp16")];
            fp16 mul_5_y_0_to_fp16 = const()[name = string("mul_5_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_5_cast_fp16 = mul(x = q_11_cast_fp16, y = mul_5_y_0_to_fp16)[name = string("mul_5_cast_fp16")];
            bool matmul_5_transpose_y_0 = const()[name = string("matmul_5_transpose_y_0"), val = bool(true)];
            bool matmul_5_transpose_x_0 = const()[name = string("matmul_5_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_5_cast_fp16, y = k_11_cast_fp16)[name = string("matmul_5_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = attn_mask_to_fp16)[name = string("add_5_cast_fp16")];
            int32 softmax_5_axis_0 = const()[name = string("softmax_5_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = add_5_cast_fp16)[name = string("softmax_5_cast_fp16")];
            bool a_11_transpose_x_0 = const()[name = string("a_11_transpose_x_0"), val = bool(false)];
            bool a_11_transpose_y_0 = const()[name = string("a_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_11_cast_fp16 = matmul(transpose_x = a_11_transpose_x_0, transpose_y = a_11_transpose_y_0, x = softmax_5_cast_fp16, y = v_11_cast_fp16)[name = string("a_11_cast_fp16")];
            tensor<int32, [1]> var_435_axes_0 = const()[name = string("op_435_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_435_cast_fp16 = squeeze(axes = var_435_axes_0, x = a_11_cast_fp16)[name = string("op_435_cast_fp16")];
            tensor<int32, [3]> var_436_perm_0 = const()[name = string("op_436_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_437 = const()[name = string("op_437"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_436_cast_fp16 = transpose(perm = var_436_perm_0, x = var_435_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [390, 896]> input_73_cast_fp16 = reshape(shape = var_437, x = var_436_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_5_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61856512))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62659392))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_5_out_bias_to_fp16 = const()[name = string("trunk_layers_5_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62661248)))];
            tensor<fp16, [390, 896]> linear_34_cast_fp16 = linear(bias = trunk_layers_5_out_bias_to_fp16, weight = trunk_layers_5_out_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [390, 896]> input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_34_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<int32, [1]> input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_5_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_5_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62663104)))];
            tensor<fp16, [896]> trunk_layers_5_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_5_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62664960)))];
            tensor<fp16, [390, 896]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = trunk_layers_5_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_5_mlp_ln_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_5_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62666816))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65878144))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_5_fc1_bias_to_fp16 = const()[name = string("trunk_layers_5_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65885376)))];
            tensor<fp16, [390, 3584]> linear_35_cast_fp16 = linear(bias = trunk_layers_5_fc1_bias_to_fp16, weight = trunk_layers_5_fc1_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_35_cast_fp16")];
            string input_79_mode_0 = const()[name = string("input_79_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = linear_35_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_5_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65892608))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69103936))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_5_fc2_bias_to_fp16 = const()[name = string("trunk_layers_5_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69105792)))];
            tensor<fp16, [390, 896]> linear_36_cast_fp16 = linear(bias = trunk_layers_5_fc2_bias_to_fp16, weight = trunk_layers_5_fc2_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<fp16, [390, 896]> input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_36_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<int32, [1]> input_83_axes_0 = const()[name = string("input_83_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_6_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_6_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69107648)))];
            tensor<fp16, [896]> trunk_layers_6_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_6_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69109504)))];
            tensor<fp16, [390, 896]> input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = trunk_layers_6_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_6_attn_ln_weight_to_fp16, x = input_81_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_6_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69111360))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69914240))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_6_q_bias_to_fp16 = const()[name = string("trunk_layers_6_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69916096)))];
            tensor<fp16, [390, 896]> linear_37_cast_fp16 = linear(bias = trunk_layers_6_q_bias_to_fp16, weight = trunk_layers_6_q_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [3]> var_471 = const()[name = string("op_471"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_472_cast_fp16 = reshape(shape = var_471, x = linear_37_cast_fp16)[name = string("op_472_cast_fp16")];
            tensor<int32, [3]> var_473_perm_0 = const()[name = string("op_473_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_13_axes_0 = const()[name = string("q_13_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_473_cast_fp16 = transpose(perm = var_473_perm_0, x = var_472_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 14, 390, 64]> q_13_cast_fp16 = expand_dims(axes = q_13_axes_0, x = var_473_cast_fp16)[name = string("q_13_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_6_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69917952))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70720832))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_6_k_bias_to_fp16 = const()[name = string("trunk_layers_6_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70722688)))];
            tensor<fp16, [390, 896]> linear_38_cast_fp16 = linear(bias = trunk_layers_6_k_bias_to_fp16, weight = trunk_layers_6_k_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [3]> var_478 = const()[name = string("op_478"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_479_cast_fp16 = reshape(shape = var_478, x = linear_38_cast_fp16)[name = string("op_479_cast_fp16")];
            tensor<int32, [3]> var_480_perm_0 = const()[name = string("op_480_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_13_axes_0 = const()[name = string("k_13_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_480_cast_fp16 = transpose(perm = var_480_perm_0, x = var_479_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 14, 390, 64]> k_13_cast_fp16 = expand_dims(axes = k_13_axes_0, x = var_480_cast_fp16)[name = string("k_13_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_6_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70724544))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71527424))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_6_v_bias_to_fp16 = const()[name = string("trunk_layers_6_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71529280)))];
            tensor<fp16, [390, 896]> linear_39_cast_fp16 = linear(bias = trunk_layers_6_v_bias_to_fp16, weight = trunk_layers_6_v_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<int32, [3]> var_485 = const()[name = string("op_485"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_486_cast_fp16 = reshape(shape = var_485, x = linear_39_cast_fp16)[name = string("op_486_cast_fp16")];
            tensor<int32, [3]> var_487_perm_0 = const()[name = string("op_487_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_13_axes_0 = const()[name = string("v_13_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_487_cast_fp16 = transpose(perm = var_487_perm_0, x = var_486_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 14, 390, 64]> v_13_cast_fp16 = expand_dims(axes = v_13_axes_0, x = var_487_cast_fp16)[name = string("v_13_cast_fp16")];
            fp16 mul_6_y_0_to_fp16 = const()[name = string("mul_6_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_6_cast_fp16 = mul(x = q_13_cast_fp16, y = mul_6_y_0_to_fp16)[name = string("mul_6_cast_fp16")];
            bool matmul_6_transpose_y_0 = const()[name = string("matmul_6_transpose_y_0"), val = bool(true)];
            bool matmul_6_transpose_x_0 = const()[name = string("matmul_6_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_6_cast_fp16, y = k_13_cast_fp16)[name = string("matmul_6_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_6_cast_fp16 = add(x = matmul_6_cast_fp16, y = attn_mask_to_fp16)[name = string("add_6_cast_fp16")];
            int32 softmax_6_axis_0 = const()[name = string("softmax_6_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = add_6_cast_fp16)[name = string("softmax_6_cast_fp16")];
            bool a_13_transpose_x_0 = const()[name = string("a_13_transpose_x_0"), val = bool(false)];
            bool a_13_transpose_y_0 = const()[name = string("a_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_13_cast_fp16 = matmul(transpose_x = a_13_transpose_x_0, transpose_y = a_13_transpose_y_0, x = softmax_6_cast_fp16, y = v_13_cast_fp16)[name = string("a_13_cast_fp16")];
            tensor<int32, [1]> var_490_axes_0 = const()[name = string("op_490_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_490_cast_fp16 = squeeze(axes = var_490_axes_0, x = a_13_cast_fp16)[name = string("op_490_cast_fp16")];
            tensor<int32, [3]> var_491_perm_0 = const()[name = string("op_491_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_492 = const()[name = string("op_492"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_491_cast_fp16 = transpose(perm = var_491_perm_0, x = var_490_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [390, 896]> input_85_cast_fp16 = reshape(shape = var_492, x = var_491_cast_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_6_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71531136))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72334016))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_6_out_bias_to_fp16 = const()[name = string("trunk_layers_6_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72335872)))];
            tensor<fp16, [390, 896]> linear_40_cast_fp16 = linear(bias = trunk_layers_6_out_bias_to_fp16, weight = trunk_layers_6_out_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [390, 896]> input_87_cast_fp16 = add(x = input_81_cast_fp16, y = linear_40_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<int32, [1]> input_89_axes_0 = const()[name = string("input_89_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_6_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_6_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72337728)))];
            tensor<fp16, [896]> trunk_layers_6_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_6_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72339584)))];
            tensor<fp16, [390, 896]> input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = trunk_layers_6_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_6_mlp_ln_weight_to_fp16, x = input_87_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_6_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72341440))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75552768))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_6_fc1_bias_to_fp16 = const()[name = string("trunk_layers_6_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75560000)))];
            tensor<fp16, [390, 3584]> linear_41_cast_fp16 = linear(bias = trunk_layers_6_fc1_bias_to_fp16, weight = trunk_layers_6_fc1_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_41_cast_fp16")];
            string input_91_mode_0 = const()[name = string("input_91_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = linear_41_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_6_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75567232))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78778560))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_6_fc2_bias_to_fp16 = const()[name = string("trunk_layers_6_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78780416)))];
            tensor<fp16, [390, 896]> linear_42_cast_fp16 = linear(bias = trunk_layers_6_fc2_bias_to_fp16, weight = trunk_layers_6_fc2_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [390, 896]> input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_42_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<int32, [1]> input_95_axes_0 = const()[name = string("input_95_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_7_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_7_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78782272)))];
            tensor<fp16, [896]> trunk_layers_7_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_7_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78784128)))];
            tensor<fp16, [390, 896]> input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = trunk_layers_7_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_7_attn_ln_weight_to_fp16, x = input_93_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_7_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78785984))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79588864))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_7_q_bias_to_fp16 = const()[name = string("trunk_layers_7_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79590720)))];
            tensor<fp16, [390, 896]> linear_43_cast_fp16 = linear(bias = trunk_layers_7_q_bias_to_fp16, weight = trunk_layers_7_q_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [3]> var_526 = const()[name = string("op_526"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_527_cast_fp16 = reshape(shape = var_526, x = linear_43_cast_fp16)[name = string("op_527_cast_fp16")];
            tensor<int32, [3]> var_528_perm_0 = const()[name = string("op_528_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_15_axes_0 = const()[name = string("q_15_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_528_cast_fp16 = transpose(perm = var_528_perm_0, x = var_527_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 14, 390, 64]> q_15_cast_fp16 = expand_dims(axes = q_15_axes_0, x = var_528_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_7_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79592576))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80395456))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_7_k_bias_to_fp16 = const()[name = string("trunk_layers_7_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80397312)))];
            tensor<fp16, [390, 896]> linear_44_cast_fp16 = linear(bias = trunk_layers_7_k_bias_to_fp16, weight = trunk_layers_7_k_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [3]> var_533 = const()[name = string("op_533"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_534_cast_fp16 = reshape(shape = var_533, x = linear_44_cast_fp16)[name = string("op_534_cast_fp16")];
            tensor<int32, [3]> var_535_perm_0 = const()[name = string("op_535_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_15_axes_0 = const()[name = string("k_15_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_535_cast_fp16 = transpose(perm = var_535_perm_0, x = var_534_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 14, 390, 64]> k_15_cast_fp16 = expand_dims(axes = k_15_axes_0, x = var_535_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_7_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80399168))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81202048))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_7_v_bias_to_fp16 = const()[name = string("trunk_layers_7_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81203904)))];
            tensor<fp16, [390, 896]> linear_45_cast_fp16 = linear(bias = trunk_layers_7_v_bias_to_fp16, weight = trunk_layers_7_v_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<int32, [3]> var_540 = const()[name = string("op_540"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_541_cast_fp16 = reshape(shape = var_540, x = linear_45_cast_fp16)[name = string("op_541_cast_fp16")];
            tensor<int32, [3]> var_542_perm_0 = const()[name = string("op_542_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_15_axes_0 = const()[name = string("v_15_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_542_cast_fp16 = transpose(perm = var_542_perm_0, x = var_541_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 14, 390, 64]> v_15_cast_fp16 = expand_dims(axes = v_15_axes_0, x = var_542_cast_fp16)[name = string("v_15_cast_fp16")];
            fp16 mul_7_y_0_to_fp16 = const()[name = string("mul_7_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_7_cast_fp16 = mul(x = q_15_cast_fp16, y = mul_7_y_0_to_fp16)[name = string("mul_7_cast_fp16")];
            bool matmul_7_transpose_y_0 = const()[name = string("matmul_7_transpose_y_0"), val = bool(true)];
            bool matmul_7_transpose_x_0 = const()[name = string("matmul_7_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_7_cast_fp16, y = k_15_cast_fp16)[name = string("matmul_7_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_7_cast_fp16 = add(x = matmul_7_cast_fp16, y = attn_mask_to_fp16)[name = string("add_7_cast_fp16")];
            int32 softmax_7_axis_0 = const()[name = string("softmax_7_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = add_7_cast_fp16)[name = string("softmax_7_cast_fp16")];
            bool a_15_transpose_x_0 = const()[name = string("a_15_transpose_x_0"), val = bool(false)];
            bool a_15_transpose_y_0 = const()[name = string("a_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_15_cast_fp16 = matmul(transpose_x = a_15_transpose_x_0, transpose_y = a_15_transpose_y_0, x = softmax_7_cast_fp16, y = v_15_cast_fp16)[name = string("a_15_cast_fp16")];
            tensor<int32, [1]> var_545_axes_0 = const()[name = string("op_545_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_545_cast_fp16 = squeeze(axes = var_545_axes_0, x = a_15_cast_fp16)[name = string("op_545_cast_fp16")];
            tensor<int32, [3]> var_546_perm_0 = const()[name = string("op_546_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_547 = const()[name = string("op_547"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_546_cast_fp16 = transpose(perm = var_546_perm_0, x = var_545_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [390, 896]> input_97_cast_fp16 = reshape(shape = var_547, x = var_546_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_7_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81205760))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82008640))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_7_out_bias_to_fp16 = const()[name = string("trunk_layers_7_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82010496)))];
            tensor<fp16, [390, 896]> linear_46_cast_fp16 = linear(bias = trunk_layers_7_out_bias_to_fp16, weight = trunk_layers_7_out_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [390, 896]> input_99_cast_fp16 = add(x = input_93_cast_fp16, y = linear_46_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<int32, [1]> input_101_axes_0 = const()[name = string("input_101_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_7_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_7_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82012352)))];
            tensor<fp16, [896]> trunk_layers_7_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_7_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82014208)))];
            tensor<fp16, [390, 896]> input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = trunk_layers_7_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_7_mlp_ln_weight_to_fp16, x = input_99_cast_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_7_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82016064))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85227392))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_7_fc1_bias_to_fp16 = const()[name = string("trunk_layers_7_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85234624)))];
            tensor<fp16, [390, 3584]> linear_47_cast_fp16 = linear(bias = trunk_layers_7_fc1_bias_to_fp16, weight = trunk_layers_7_fc1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_47_cast_fp16")];
            string input_103_mode_0 = const()[name = string("input_103_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_103_cast_fp16 = gelu(mode = input_103_mode_0, x = linear_47_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_7_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85241856))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88453184))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_7_fc2_bias_to_fp16 = const()[name = string("trunk_layers_7_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88455040)))];
            tensor<fp16, [390, 896]> linear_48_cast_fp16 = linear(bias = trunk_layers_7_fc2_bias_to_fp16, weight = trunk_layers_7_fc2_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [390, 896]> input_105_cast_fp16 = add(x = input_99_cast_fp16, y = linear_48_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<int32, [1]> input_107_axes_0 = const()[name = string("input_107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_8_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_8_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88456896)))];
            tensor<fp16, [896]> trunk_layers_8_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_8_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88458752)))];
            tensor<fp16, [390, 896]> input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = trunk_layers_8_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_8_attn_ln_weight_to_fp16, x = input_105_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_8_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88460608))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89263488))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_8_q_bias_to_fp16 = const()[name = string("trunk_layers_8_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89265344)))];
            tensor<fp16, [390, 896]> linear_49_cast_fp16 = linear(bias = trunk_layers_8_q_bias_to_fp16, weight = trunk_layers_8_q_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [3]> var_581 = const()[name = string("op_581"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_582_cast_fp16 = reshape(shape = var_581, x = linear_49_cast_fp16)[name = string("op_582_cast_fp16")];
            tensor<int32, [3]> var_583_perm_0 = const()[name = string("op_583_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_17_axes_0 = const()[name = string("q_17_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_583_cast_fp16 = transpose(perm = var_583_perm_0, x = var_582_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 14, 390, 64]> q_17_cast_fp16 = expand_dims(axes = q_17_axes_0, x = var_583_cast_fp16)[name = string("q_17_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_8_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89267200))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90070080))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_8_k_bias_to_fp16 = const()[name = string("trunk_layers_8_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90071936)))];
            tensor<fp16, [390, 896]> linear_50_cast_fp16 = linear(bias = trunk_layers_8_k_bias_to_fp16, weight = trunk_layers_8_k_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [3]> var_588 = const()[name = string("op_588"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_589_cast_fp16 = reshape(shape = var_588, x = linear_50_cast_fp16)[name = string("op_589_cast_fp16")];
            tensor<int32, [3]> var_590_perm_0 = const()[name = string("op_590_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_17_axes_0 = const()[name = string("k_17_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_590_cast_fp16 = transpose(perm = var_590_perm_0, x = var_589_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 14, 390, 64]> k_17_cast_fp16 = expand_dims(axes = k_17_axes_0, x = var_590_cast_fp16)[name = string("k_17_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_8_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90073792))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90876672))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_8_v_bias_to_fp16 = const()[name = string("trunk_layers_8_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90878528)))];
            tensor<fp16, [390, 896]> linear_51_cast_fp16 = linear(bias = trunk_layers_8_v_bias_to_fp16, weight = trunk_layers_8_v_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [3]> var_595 = const()[name = string("op_595"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_596_cast_fp16 = reshape(shape = var_595, x = linear_51_cast_fp16)[name = string("op_596_cast_fp16")];
            tensor<int32, [3]> var_597_perm_0 = const()[name = string("op_597_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_17_axes_0 = const()[name = string("v_17_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_597_cast_fp16 = transpose(perm = var_597_perm_0, x = var_596_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 14, 390, 64]> v_17_cast_fp16 = expand_dims(axes = v_17_axes_0, x = var_597_cast_fp16)[name = string("v_17_cast_fp16")];
            fp16 mul_8_y_0_to_fp16 = const()[name = string("mul_8_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_8_cast_fp16 = mul(x = q_17_cast_fp16, y = mul_8_y_0_to_fp16)[name = string("mul_8_cast_fp16")];
            bool matmul_8_transpose_y_0 = const()[name = string("matmul_8_transpose_y_0"), val = bool(true)];
            bool matmul_8_transpose_x_0 = const()[name = string("matmul_8_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_8_cast_fp16, y = k_17_cast_fp16)[name = string("matmul_8_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_8_cast_fp16 = add(x = matmul_8_cast_fp16, y = attn_mask_to_fp16)[name = string("add_8_cast_fp16")];
            int32 softmax_8_axis_0 = const()[name = string("softmax_8_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = add_8_cast_fp16)[name = string("softmax_8_cast_fp16")];
            bool a_17_transpose_x_0 = const()[name = string("a_17_transpose_x_0"), val = bool(false)];
            bool a_17_transpose_y_0 = const()[name = string("a_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_17_cast_fp16 = matmul(transpose_x = a_17_transpose_x_0, transpose_y = a_17_transpose_y_0, x = softmax_8_cast_fp16, y = v_17_cast_fp16)[name = string("a_17_cast_fp16")];
            tensor<int32, [1]> var_600_axes_0 = const()[name = string("op_600_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_600_cast_fp16 = squeeze(axes = var_600_axes_0, x = a_17_cast_fp16)[name = string("op_600_cast_fp16")];
            tensor<int32, [3]> var_601_perm_0 = const()[name = string("op_601_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_602 = const()[name = string("op_602"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_601_cast_fp16 = transpose(perm = var_601_perm_0, x = var_600_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [390, 896]> input_109_cast_fp16 = reshape(shape = var_602, x = var_601_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_8_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90880384))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91683264))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_8_out_bias_to_fp16 = const()[name = string("trunk_layers_8_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91685120)))];
            tensor<fp16, [390, 896]> linear_52_cast_fp16 = linear(bias = trunk_layers_8_out_bias_to_fp16, weight = trunk_layers_8_out_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [390, 896]> input_111_cast_fp16 = add(x = input_105_cast_fp16, y = linear_52_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<int32, [1]> input_113_axes_0 = const()[name = string("input_113_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_8_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_8_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91686976)))];
            tensor<fp16, [896]> trunk_layers_8_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_8_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91688832)))];
            tensor<fp16, [390, 896]> input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = trunk_layers_8_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_8_mlp_ln_weight_to_fp16, x = input_111_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_8_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91690688))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94902016))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_8_fc1_bias_to_fp16 = const()[name = string("trunk_layers_8_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94909248)))];
            tensor<fp16, [390, 3584]> linear_53_cast_fp16 = linear(bias = trunk_layers_8_fc1_bias_to_fp16, weight = trunk_layers_8_fc1_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_53_cast_fp16")];
            string input_115_mode_0 = const()[name = string("input_115_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_115_cast_fp16 = gelu(mode = input_115_mode_0, x = linear_53_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_8_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94916480))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98127808))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_8_fc2_bias_to_fp16 = const()[name = string("trunk_layers_8_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98129664)))];
            tensor<fp16, [390, 896]> linear_54_cast_fp16 = linear(bias = trunk_layers_8_fc2_bias_to_fp16, weight = trunk_layers_8_fc2_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [390, 896]> input_117_cast_fp16 = add(x = input_111_cast_fp16, y = linear_54_cast_fp16)[name = string("input_117_cast_fp16")];
            tensor<int32, [1]> input_119_axes_0 = const()[name = string("input_119_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_9_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_9_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98131520)))];
            tensor<fp16, [896]> trunk_layers_9_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_9_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98133376)))];
            tensor<fp16, [390, 896]> input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = trunk_layers_9_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_9_attn_ln_weight_to_fp16, x = input_117_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_9_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98135232))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98938112))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_9_q_bias_to_fp16 = const()[name = string("trunk_layers_9_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98939968)))];
            tensor<fp16, [390, 896]> linear_55_cast_fp16 = linear(bias = trunk_layers_9_q_bias_to_fp16, weight = trunk_layers_9_q_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<int32, [3]> var_636 = const()[name = string("op_636"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_637_cast_fp16 = reshape(shape = var_636, x = linear_55_cast_fp16)[name = string("op_637_cast_fp16")];
            tensor<int32, [3]> var_638_perm_0 = const()[name = string("op_638_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_19_axes_0 = const()[name = string("q_19_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_638_cast_fp16 = transpose(perm = var_638_perm_0, x = var_637_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 14, 390, 64]> q_19_cast_fp16 = expand_dims(axes = q_19_axes_0, x = var_638_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_9_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98941824))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99744704))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_9_k_bias_to_fp16 = const()[name = string("trunk_layers_9_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99746560)))];
            tensor<fp16, [390, 896]> linear_56_cast_fp16 = linear(bias = trunk_layers_9_k_bias_to_fp16, weight = trunk_layers_9_k_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [3]> var_643 = const()[name = string("op_643"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_644_cast_fp16 = reshape(shape = var_643, x = linear_56_cast_fp16)[name = string("op_644_cast_fp16")];
            tensor<int32, [3]> var_645_perm_0 = const()[name = string("op_645_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_19_axes_0 = const()[name = string("k_19_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_645_cast_fp16 = transpose(perm = var_645_perm_0, x = var_644_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 14, 390, 64]> k_19_cast_fp16 = expand_dims(axes = k_19_axes_0, x = var_645_cast_fp16)[name = string("k_19_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_9_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99748416))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100551296))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_9_v_bias_to_fp16 = const()[name = string("trunk_layers_9_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100553152)))];
            tensor<fp16, [390, 896]> linear_57_cast_fp16 = linear(bias = trunk_layers_9_v_bias_to_fp16, weight = trunk_layers_9_v_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [3]> var_650 = const()[name = string("op_650"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_651_cast_fp16 = reshape(shape = var_650, x = linear_57_cast_fp16)[name = string("op_651_cast_fp16")];
            tensor<int32, [3]> var_652_perm_0 = const()[name = string("op_652_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_19_axes_0 = const()[name = string("v_19_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_652_cast_fp16 = transpose(perm = var_652_perm_0, x = var_651_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 14, 390, 64]> v_19_cast_fp16 = expand_dims(axes = v_19_axes_0, x = var_652_cast_fp16)[name = string("v_19_cast_fp16")];
            fp16 mul_9_y_0_to_fp16 = const()[name = string("mul_9_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_9_cast_fp16 = mul(x = q_19_cast_fp16, y = mul_9_y_0_to_fp16)[name = string("mul_9_cast_fp16")];
            bool matmul_9_transpose_y_0 = const()[name = string("matmul_9_transpose_y_0"), val = bool(true)];
            bool matmul_9_transpose_x_0 = const()[name = string("matmul_9_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_9_cast_fp16, y = k_19_cast_fp16)[name = string("matmul_9_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_9_cast_fp16 = add(x = matmul_9_cast_fp16, y = attn_mask_to_fp16)[name = string("add_9_cast_fp16")];
            int32 softmax_9_axis_0 = const()[name = string("softmax_9_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = add_9_cast_fp16)[name = string("softmax_9_cast_fp16")];
            bool a_19_transpose_x_0 = const()[name = string("a_19_transpose_x_0"), val = bool(false)];
            bool a_19_transpose_y_0 = const()[name = string("a_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_19_cast_fp16 = matmul(transpose_x = a_19_transpose_x_0, transpose_y = a_19_transpose_y_0, x = softmax_9_cast_fp16, y = v_19_cast_fp16)[name = string("a_19_cast_fp16")];
            tensor<int32, [1]> var_655_axes_0 = const()[name = string("op_655_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_655_cast_fp16 = squeeze(axes = var_655_axes_0, x = a_19_cast_fp16)[name = string("op_655_cast_fp16")];
            tensor<int32, [3]> var_656_perm_0 = const()[name = string("op_656_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_657 = const()[name = string("op_657"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_656_cast_fp16 = transpose(perm = var_656_perm_0, x = var_655_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [390, 896]> input_121_cast_fp16 = reshape(shape = var_657, x = var_656_cast_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_9_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100555008))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101357888))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_9_out_bias_to_fp16 = const()[name = string("trunk_layers_9_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101359744)))];
            tensor<fp16, [390, 896]> linear_58_cast_fp16 = linear(bias = trunk_layers_9_out_bias_to_fp16, weight = trunk_layers_9_out_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<fp16, [390, 896]> input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_58_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<int32, [1]> input_125_axes_0 = const()[name = string("input_125_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_9_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_9_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101361600)))];
            tensor<fp16, [896]> trunk_layers_9_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_9_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101363456)))];
            tensor<fp16, [390, 896]> input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = trunk_layers_9_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_9_mlp_ln_weight_to_fp16, x = input_123_cast_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_9_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101365312))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104576640))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_9_fc1_bias_to_fp16 = const()[name = string("trunk_layers_9_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104583872)))];
            tensor<fp16, [390, 3584]> linear_59_cast_fp16 = linear(bias = trunk_layers_9_fc1_bias_to_fp16, weight = trunk_layers_9_fc1_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_59_cast_fp16")];
            string input_127_mode_0 = const()[name = string("input_127_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_127_cast_fp16 = gelu(mode = input_127_mode_0, x = linear_59_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_9_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104591104))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107802432))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_9_fc2_bias_to_fp16 = const()[name = string("trunk_layers_9_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107804288)))];
            tensor<fp16, [390, 896]> linear_60_cast_fp16 = linear(bias = trunk_layers_9_fc2_bias_to_fp16, weight = trunk_layers_9_fc2_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [390, 896]> input_129_cast_fp16 = add(x = input_123_cast_fp16, y = linear_60_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<int32, [1]> input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_10_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_10_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107806144)))];
            tensor<fp16, [896]> trunk_layers_10_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_10_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107808000)))];
            tensor<fp16, [390, 896]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = trunk_layers_10_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_10_attn_ln_weight_to_fp16, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_10_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107809856))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108612736))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_10_q_bias_to_fp16 = const()[name = string("trunk_layers_10_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108614592)))];
            tensor<fp16, [390, 896]> linear_61_cast_fp16 = linear(bias = trunk_layers_10_q_bias_to_fp16, weight = trunk_layers_10_q_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<int32, [3]> var_691 = const()[name = string("op_691"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_692_cast_fp16 = reshape(shape = var_691, x = linear_61_cast_fp16)[name = string("op_692_cast_fp16")];
            tensor<int32, [3]> var_693_perm_0 = const()[name = string("op_693_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_21_axes_0 = const()[name = string("q_21_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_693_cast_fp16 = transpose(perm = var_693_perm_0, x = var_692_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 14, 390, 64]> q_21_cast_fp16 = expand_dims(axes = q_21_axes_0, x = var_693_cast_fp16)[name = string("q_21_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_10_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108616448))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109419328))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_10_k_bias_to_fp16 = const()[name = string("trunk_layers_10_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109421184)))];
            tensor<fp16, [390, 896]> linear_62_cast_fp16 = linear(bias = trunk_layers_10_k_bias_to_fp16, weight = trunk_layers_10_k_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<int32, [3]> var_698 = const()[name = string("op_698"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_699_cast_fp16 = reshape(shape = var_698, x = linear_62_cast_fp16)[name = string("op_699_cast_fp16")];
            tensor<int32, [3]> var_700_perm_0 = const()[name = string("op_700_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_21_axes_0 = const()[name = string("k_21_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_700_cast_fp16 = transpose(perm = var_700_perm_0, x = var_699_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 14, 390, 64]> k_21_cast_fp16 = expand_dims(axes = k_21_axes_0, x = var_700_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_10_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109423040))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110225920))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_10_v_bias_to_fp16 = const()[name = string("trunk_layers_10_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110227776)))];
            tensor<fp16, [390, 896]> linear_63_cast_fp16 = linear(bias = trunk_layers_10_v_bias_to_fp16, weight = trunk_layers_10_v_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [3]> var_705 = const()[name = string("op_705"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_706_cast_fp16 = reshape(shape = var_705, x = linear_63_cast_fp16)[name = string("op_706_cast_fp16")];
            tensor<int32, [3]> var_707_perm_0 = const()[name = string("op_707_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_21_axes_0 = const()[name = string("v_21_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_707_cast_fp16 = transpose(perm = var_707_perm_0, x = var_706_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 14, 390, 64]> v_21_cast_fp16 = expand_dims(axes = v_21_axes_0, x = var_707_cast_fp16)[name = string("v_21_cast_fp16")];
            fp16 mul_10_y_0_to_fp16 = const()[name = string("mul_10_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_10_cast_fp16 = mul(x = q_21_cast_fp16, y = mul_10_y_0_to_fp16)[name = string("mul_10_cast_fp16")];
            bool matmul_10_transpose_y_0 = const()[name = string("matmul_10_transpose_y_0"), val = bool(true)];
            bool matmul_10_transpose_x_0 = const()[name = string("matmul_10_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_10_cast_fp16, y = k_21_cast_fp16)[name = string("matmul_10_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_10_cast_fp16 = add(x = matmul_10_cast_fp16, y = attn_mask_to_fp16)[name = string("add_10_cast_fp16")];
            int32 softmax_10_axis_0 = const()[name = string("softmax_10_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = add_10_cast_fp16)[name = string("softmax_10_cast_fp16")];
            bool a_21_transpose_x_0 = const()[name = string("a_21_transpose_x_0"), val = bool(false)];
            bool a_21_transpose_y_0 = const()[name = string("a_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_21_cast_fp16 = matmul(transpose_x = a_21_transpose_x_0, transpose_y = a_21_transpose_y_0, x = softmax_10_cast_fp16, y = v_21_cast_fp16)[name = string("a_21_cast_fp16")];
            tensor<int32, [1]> var_710_axes_0 = const()[name = string("op_710_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_710_cast_fp16 = squeeze(axes = var_710_axes_0, x = a_21_cast_fp16)[name = string("op_710_cast_fp16")];
            tensor<int32, [3]> var_711_perm_0 = const()[name = string("op_711_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_712 = const()[name = string("op_712"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_711_cast_fp16 = transpose(perm = var_711_perm_0, x = var_710_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [390, 896]> input_133_cast_fp16 = reshape(shape = var_712, x = var_711_cast_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_10_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110229632))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111032512))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_10_out_bias_to_fp16 = const()[name = string("trunk_layers_10_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111034368)))];
            tensor<fp16, [390, 896]> linear_64_cast_fp16 = linear(bias = trunk_layers_10_out_bias_to_fp16, weight = trunk_layers_10_out_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<fp16, [390, 896]> input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_64_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<int32, [1]> input_137_axes_0 = const()[name = string("input_137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_10_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_10_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111036224)))];
            tensor<fp16, [896]> trunk_layers_10_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_10_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111038080)))];
            tensor<fp16, [390, 896]> input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = trunk_layers_10_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_10_mlp_ln_weight_to_fp16, x = input_135_cast_fp16)[name = string("input_137_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_10_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111039936))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114251264))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_10_fc1_bias_to_fp16 = const()[name = string("trunk_layers_10_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114258496)))];
            tensor<fp16, [390, 3584]> linear_65_cast_fp16 = linear(bias = trunk_layers_10_fc1_bias_to_fp16, weight = trunk_layers_10_fc1_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_65_cast_fp16")];
            string input_139_mode_0 = const()[name = string("input_139_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_139_cast_fp16 = gelu(mode = input_139_mode_0, x = linear_65_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_10_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114265728))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117477056))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_10_fc2_bias_to_fp16 = const()[name = string("trunk_layers_10_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117478912)))];
            tensor<fp16, [390, 896]> linear_66_cast_fp16 = linear(bias = trunk_layers_10_fc2_bias_to_fp16, weight = trunk_layers_10_fc2_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [390, 896]> input_141_cast_fp16 = add(x = input_135_cast_fp16, y = linear_66_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<int32, [1]> input_143_axes_0 = const()[name = string("input_143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_11_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_11_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117480768)))];
            tensor<fp16, [896]> trunk_layers_11_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_11_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117482624)))];
            tensor<fp16, [390, 896]> input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = trunk_layers_11_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_11_attn_ln_weight_to_fp16, x = input_141_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_11_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117484480))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118287360))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_11_q_bias_to_fp16 = const()[name = string("trunk_layers_11_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118289216)))];
            tensor<fp16, [390, 896]> linear_67_cast_fp16 = linear(bias = trunk_layers_11_q_bias_to_fp16, weight = trunk_layers_11_q_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<int32, [3]> var_746 = const()[name = string("op_746"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_747_cast_fp16 = reshape(shape = var_746, x = linear_67_cast_fp16)[name = string("op_747_cast_fp16")];
            tensor<int32, [3]> var_748_perm_0 = const()[name = string("op_748_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_23_axes_0 = const()[name = string("q_23_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_748_cast_fp16 = transpose(perm = var_748_perm_0, x = var_747_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 14, 390, 64]> q_23_cast_fp16 = expand_dims(axes = q_23_axes_0, x = var_748_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_11_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118291072))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119093952))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_11_k_bias_to_fp16 = const()[name = string("trunk_layers_11_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119095808)))];
            tensor<fp16, [390, 896]> linear_68_cast_fp16 = linear(bias = trunk_layers_11_k_bias_to_fp16, weight = trunk_layers_11_k_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<int32, [3]> var_753 = const()[name = string("op_753"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_754_cast_fp16 = reshape(shape = var_753, x = linear_68_cast_fp16)[name = string("op_754_cast_fp16")];
            tensor<int32, [3]> var_755_perm_0 = const()[name = string("op_755_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_23_axes_0 = const()[name = string("k_23_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_755_cast_fp16 = transpose(perm = var_755_perm_0, x = var_754_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 14, 390, 64]> k_23_cast_fp16 = expand_dims(axes = k_23_axes_0, x = var_755_cast_fp16)[name = string("k_23_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_11_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119097664))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119900544))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_11_v_bias_to_fp16 = const()[name = string("trunk_layers_11_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119902400)))];
            tensor<fp16, [390, 896]> linear_69_cast_fp16 = linear(bias = trunk_layers_11_v_bias_to_fp16, weight = trunk_layers_11_v_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<int32, [3]> var_760 = const()[name = string("op_760"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_761_cast_fp16 = reshape(shape = var_760, x = linear_69_cast_fp16)[name = string("op_761_cast_fp16")];
            tensor<int32, [3]> var_762_perm_0 = const()[name = string("op_762_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_23_axes_0 = const()[name = string("v_23_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_762_cast_fp16 = transpose(perm = var_762_perm_0, x = var_761_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 14, 390, 64]> v_23_cast_fp16 = expand_dims(axes = v_23_axes_0, x = var_762_cast_fp16)[name = string("v_23_cast_fp16")];
            fp16 mul_11_y_0_to_fp16 = const()[name = string("mul_11_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_11_cast_fp16 = mul(x = q_23_cast_fp16, y = mul_11_y_0_to_fp16)[name = string("mul_11_cast_fp16")];
            bool matmul_11_transpose_y_0 = const()[name = string("matmul_11_transpose_y_0"), val = bool(true)];
            bool matmul_11_transpose_x_0 = const()[name = string("matmul_11_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_11_cast_fp16, y = k_23_cast_fp16)[name = string("matmul_11_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_11_cast_fp16 = add(x = matmul_11_cast_fp16, y = attn_mask_to_fp16)[name = string("add_11_cast_fp16")];
            int32 softmax_11_axis_0 = const()[name = string("softmax_11_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = add_11_cast_fp16)[name = string("softmax_11_cast_fp16")];
            bool a_23_transpose_x_0 = const()[name = string("a_23_transpose_x_0"), val = bool(false)];
            bool a_23_transpose_y_0 = const()[name = string("a_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_23_cast_fp16 = matmul(transpose_x = a_23_transpose_x_0, transpose_y = a_23_transpose_y_0, x = softmax_11_cast_fp16, y = v_23_cast_fp16)[name = string("a_23_cast_fp16")];
            tensor<int32, [1]> var_765_axes_0 = const()[name = string("op_765_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_765_cast_fp16 = squeeze(axes = var_765_axes_0, x = a_23_cast_fp16)[name = string("op_765_cast_fp16")];
            tensor<int32, [3]> var_766_perm_0 = const()[name = string("op_766_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_767 = const()[name = string("op_767"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_766_cast_fp16 = transpose(perm = var_766_perm_0, x = var_765_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [390, 896]> input_145_cast_fp16 = reshape(shape = var_767, x = var_766_cast_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_11_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119904256))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120707136))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_11_out_bias_to_fp16 = const()[name = string("trunk_layers_11_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120708992)))];
            tensor<fp16, [390, 896]> linear_70_cast_fp16 = linear(bias = trunk_layers_11_out_bias_to_fp16, weight = trunk_layers_11_out_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [390, 896]> input_147_cast_fp16 = add(x = input_141_cast_fp16, y = linear_70_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<int32, [1]> input_149_axes_0 = const()[name = string("input_149_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_11_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_11_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120710848)))];
            tensor<fp16, [896]> trunk_layers_11_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_11_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120712704)))];
            tensor<fp16, [390, 896]> input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = trunk_layers_11_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_11_mlp_ln_weight_to_fp16, x = input_147_cast_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_11_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120714560))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123925888))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_11_fc1_bias_to_fp16 = const()[name = string("trunk_layers_11_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123933120)))];
            tensor<fp16, [390, 3584]> linear_71_cast_fp16 = linear(bias = trunk_layers_11_fc1_bias_to_fp16, weight = trunk_layers_11_fc1_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_71_cast_fp16")];
            string input_151_mode_0 = const()[name = string("input_151_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = linear_71_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_11_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123940352))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127151680))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_11_fc2_bias_to_fp16 = const()[name = string("trunk_layers_11_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127153536)))];
            tensor<fp16, [390, 896]> linear_72_cast_fp16 = linear(bias = trunk_layers_11_fc2_bias_to_fp16, weight = trunk_layers_11_fc2_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<fp16, [390, 896]> input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_72_cast_fp16)[name = string("input_153_cast_fp16")];
            tensor<int32, [1]> input_155_axes_0 = const()[name = string("input_155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_12_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_12_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127155392)))];
            tensor<fp16, [896]> trunk_layers_12_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_12_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127157248)))];
            tensor<fp16, [390, 896]> input_155_cast_fp16 = layer_norm(axes = input_155_axes_0, beta = trunk_layers_12_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_12_attn_ln_weight_to_fp16, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_12_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127159104))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127961984))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_12_q_bias_to_fp16 = const()[name = string("trunk_layers_12_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127963840)))];
            tensor<fp16, [390, 896]> linear_73_cast_fp16 = linear(bias = trunk_layers_12_q_bias_to_fp16, weight = trunk_layers_12_q_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<int32, [3]> var_801 = const()[name = string("op_801"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_802_cast_fp16 = reshape(shape = var_801, x = linear_73_cast_fp16)[name = string("op_802_cast_fp16")];
            tensor<int32, [3]> var_803_perm_0 = const()[name = string("op_803_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_25_axes_0 = const()[name = string("q_25_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_803_cast_fp16 = transpose(perm = var_803_perm_0, x = var_802_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 14, 390, 64]> q_25_cast_fp16 = expand_dims(axes = q_25_axes_0, x = var_803_cast_fp16)[name = string("q_25_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_12_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127965696))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128768576))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_12_k_bias_to_fp16 = const()[name = string("trunk_layers_12_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128770432)))];
            tensor<fp16, [390, 896]> linear_74_cast_fp16 = linear(bias = trunk_layers_12_k_bias_to_fp16, weight = trunk_layers_12_k_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<int32, [3]> var_808 = const()[name = string("op_808"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_809_cast_fp16 = reshape(shape = var_808, x = linear_74_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<int32, [3]> var_810_perm_0 = const()[name = string("op_810_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_25_axes_0 = const()[name = string("k_25_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_810_cast_fp16 = transpose(perm = var_810_perm_0, x = var_809_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 14, 390, 64]> k_25_cast_fp16 = expand_dims(axes = k_25_axes_0, x = var_810_cast_fp16)[name = string("k_25_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_12_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128772288))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129575168))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_12_v_bias_to_fp16 = const()[name = string("trunk_layers_12_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129577024)))];
            tensor<fp16, [390, 896]> linear_75_cast_fp16 = linear(bias = trunk_layers_12_v_bias_to_fp16, weight = trunk_layers_12_v_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<int32, [3]> var_815 = const()[name = string("op_815"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_816_cast_fp16 = reshape(shape = var_815, x = linear_75_cast_fp16)[name = string("op_816_cast_fp16")];
            tensor<int32, [3]> var_817_perm_0 = const()[name = string("op_817_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_25_axes_0 = const()[name = string("v_25_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_817_cast_fp16 = transpose(perm = var_817_perm_0, x = var_816_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 14, 390, 64]> v_25_cast_fp16 = expand_dims(axes = v_25_axes_0, x = var_817_cast_fp16)[name = string("v_25_cast_fp16")];
            fp16 mul_12_y_0_to_fp16 = const()[name = string("mul_12_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_12_cast_fp16 = mul(x = q_25_cast_fp16, y = mul_12_y_0_to_fp16)[name = string("mul_12_cast_fp16")];
            bool matmul_12_transpose_y_0 = const()[name = string("matmul_12_transpose_y_0"), val = bool(true)];
            bool matmul_12_transpose_x_0 = const()[name = string("matmul_12_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = mul_12_cast_fp16, y = k_25_cast_fp16)[name = string("matmul_12_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_12_cast_fp16 = add(x = matmul_12_cast_fp16, y = attn_mask_to_fp16)[name = string("add_12_cast_fp16")];
            int32 softmax_12_axis_0 = const()[name = string("softmax_12_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = add_12_cast_fp16)[name = string("softmax_12_cast_fp16")];
            bool a_25_transpose_x_0 = const()[name = string("a_25_transpose_x_0"), val = bool(false)];
            bool a_25_transpose_y_0 = const()[name = string("a_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_25_cast_fp16 = matmul(transpose_x = a_25_transpose_x_0, transpose_y = a_25_transpose_y_0, x = softmax_12_cast_fp16, y = v_25_cast_fp16)[name = string("a_25_cast_fp16")];
            tensor<int32, [1]> var_820_axes_0 = const()[name = string("op_820_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_820_cast_fp16 = squeeze(axes = var_820_axes_0, x = a_25_cast_fp16)[name = string("op_820_cast_fp16")];
            tensor<int32, [3]> var_821_perm_0 = const()[name = string("op_821_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_822 = const()[name = string("op_822"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_821_cast_fp16 = transpose(perm = var_821_perm_0, x = var_820_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [390, 896]> input_157_cast_fp16 = reshape(shape = var_822, x = var_821_cast_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_12_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129578880))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130381760))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_12_out_bias_to_fp16 = const()[name = string("trunk_layers_12_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130383616)))];
            tensor<fp16, [390, 896]> linear_76_cast_fp16 = linear(bias = trunk_layers_12_out_bias_to_fp16, weight = trunk_layers_12_out_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [390, 896]> input_159_cast_fp16 = add(x = input_153_cast_fp16, y = linear_76_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<int32, [1]> input_161_axes_0 = const()[name = string("input_161_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_12_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_12_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130385472)))];
            tensor<fp16, [896]> trunk_layers_12_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_12_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130387328)))];
            tensor<fp16, [390, 896]> input_161_cast_fp16 = layer_norm(axes = input_161_axes_0, beta = trunk_layers_12_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_12_mlp_ln_weight_to_fp16, x = input_159_cast_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_12_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130389184))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133600512))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_12_fc1_bias_to_fp16 = const()[name = string("trunk_layers_12_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133607744)))];
            tensor<fp16, [390, 3584]> linear_77_cast_fp16 = linear(bias = trunk_layers_12_fc1_bias_to_fp16, weight = trunk_layers_12_fc1_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_77_cast_fp16")];
            string input_163_mode_0 = const()[name = string("input_163_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_163_cast_fp16 = gelu(mode = input_163_mode_0, x = linear_77_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_12_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133614976))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136826304))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_12_fc2_bias_to_fp16 = const()[name = string("trunk_layers_12_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136828160)))];
            tensor<fp16, [390, 896]> linear_78_cast_fp16 = linear(bias = trunk_layers_12_fc2_bias_to_fp16, weight = trunk_layers_12_fc2_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<fp16, [390, 896]> input_165_cast_fp16 = add(x = input_159_cast_fp16, y = linear_78_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<int32, [1]> input_167_axes_0 = const()[name = string("input_167_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_13_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_13_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136830016)))];
            tensor<fp16, [896]> trunk_layers_13_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_13_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136831872)))];
            tensor<fp16, [390, 896]> input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = trunk_layers_13_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_13_attn_ln_weight_to_fp16, x = input_165_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_13_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136833728))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137636608))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_13_q_bias_to_fp16 = const()[name = string("trunk_layers_13_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137638464)))];
            tensor<fp16, [390, 896]> linear_79_cast_fp16 = linear(bias = trunk_layers_13_q_bias_to_fp16, weight = trunk_layers_13_q_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [3]> var_856 = const()[name = string("op_856"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_857_cast_fp16 = reshape(shape = var_856, x = linear_79_cast_fp16)[name = string("op_857_cast_fp16")];
            tensor<int32, [3]> var_858_perm_0 = const()[name = string("op_858_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_27_axes_0 = const()[name = string("q_27_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_858_cast_fp16 = transpose(perm = var_858_perm_0, x = var_857_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 14, 390, 64]> q_27_cast_fp16 = expand_dims(axes = q_27_axes_0, x = var_858_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_13_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137640320))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138443200))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_13_k_bias_to_fp16 = const()[name = string("trunk_layers_13_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138445056)))];
            tensor<fp16, [390, 896]> linear_80_cast_fp16 = linear(bias = trunk_layers_13_k_bias_to_fp16, weight = trunk_layers_13_k_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<int32, [3]> var_863 = const()[name = string("op_863"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_864_cast_fp16 = reshape(shape = var_863, x = linear_80_cast_fp16)[name = string("op_864_cast_fp16")];
            tensor<int32, [3]> var_865_perm_0 = const()[name = string("op_865_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_27_axes_0 = const()[name = string("k_27_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_865_cast_fp16 = transpose(perm = var_865_perm_0, x = var_864_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 14, 390, 64]> k_27_cast_fp16 = expand_dims(axes = k_27_axes_0, x = var_865_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_13_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138446912))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139249792))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_13_v_bias_to_fp16 = const()[name = string("trunk_layers_13_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139251648)))];
            tensor<fp16, [390, 896]> linear_81_cast_fp16 = linear(bias = trunk_layers_13_v_bias_to_fp16, weight = trunk_layers_13_v_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<int32, [3]> var_870 = const()[name = string("op_870"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_871_cast_fp16 = reshape(shape = var_870, x = linear_81_cast_fp16)[name = string("op_871_cast_fp16")];
            tensor<int32, [3]> var_872_perm_0 = const()[name = string("op_872_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_27_axes_0 = const()[name = string("v_27_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_872_cast_fp16 = transpose(perm = var_872_perm_0, x = var_871_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 14, 390, 64]> v_27_cast_fp16 = expand_dims(axes = v_27_axes_0, x = var_872_cast_fp16)[name = string("v_27_cast_fp16")];
            fp16 mul_13_y_0_to_fp16 = const()[name = string("mul_13_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_13_cast_fp16 = mul(x = q_27_cast_fp16, y = mul_13_y_0_to_fp16)[name = string("mul_13_cast_fp16")];
            bool matmul_13_transpose_y_0 = const()[name = string("matmul_13_transpose_y_0"), val = bool(true)];
            bool matmul_13_transpose_x_0 = const()[name = string("matmul_13_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = mul_13_cast_fp16, y = k_27_cast_fp16)[name = string("matmul_13_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_13_cast_fp16 = add(x = matmul_13_cast_fp16, y = attn_mask_to_fp16)[name = string("add_13_cast_fp16")];
            int32 softmax_13_axis_0 = const()[name = string("softmax_13_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = add_13_cast_fp16)[name = string("softmax_13_cast_fp16")];
            bool a_27_transpose_x_0 = const()[name = string("a_27_transpose_x_0"), val = bool(false)];
            bool a_27_transpose_y_0 = const()[name = string("a_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_27_cast_fp16 = matmul(transpose_x = a_27_transpose_x_0, transpose_y = a_27_transpose_y_0, x = softmax_13_cast_fp16, y = v_27_cast_fp16)[name = string("a_27_cast_fp16")];
            tensor<int32, [1]> var_875_axes_0 = const()[name = string("op_875_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_875_cast_fp16 = squeeze(axes = var_875_axes_0, x = a_27_cast_fp16)[name = string("op_875_cast_fp16")];
            tensor<int32, [3]> var_876_perm_0 = const()[name = string("op_876_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_877 = const()[name = string("op_877"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_876_cast_fp16 = transpose(perm = var_876_perm_0, x = var_875_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [390, 896]> input_169_cast_fp16 = reshape(shape = var_877, x = var_876_cast_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_13_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139253504))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140056384))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_13_out_bias_to_fp16 = const()[name = string("trunk_layers_13_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140058240)))];
            tensor<fp16, [390, 896]> linear_82_cast_fp16 = linear(bias = trunk_layers_13_out_bias_to_fp16, weight = trunk_layers_13_out_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [390, 896]> input_171_cast_fp16 = add(x = input_165_cast_fp16, y = linear_82_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<int32, [1]> input_173_axes_0 = const()[name = string("input_173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_13_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_13_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140060096)))];
            tensor<fp16, [896]> trunk_layers_13_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_13_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140061952)))];
            tensor<fp16, [390, 896]> input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = trunk_layers_13_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_13_mlp_ln_weight_to_fp16, x = input_171_cast_fp16)[name = string("input_173_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_13_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140063808))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143275136))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_13_fc1_bias_to_fp16 = const()[name = string("trunk_layers_13_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143282368)))];
            tensor<fp16, [390, 3584]> linear_83_cast_fp16 = linear(bias = trunk_layers_13_fc1_bias_to_fp16, weight = trunk_layers_13_fc1_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_83_cast_fp16")];
            string input_175_mode_0 = const()[name = string("input_175_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_175_cast_fp16 = gelu(mode = input_175_mode_0, x = linear_83_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_13_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143289600))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146500928))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_13_fc2_bias_to_fp16 = const()[name = string("trunk_layers_13_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146502784)))];
            tensor<fp16, [390, 896]> linear_84_cast_fp16 = linear(bias = trunk_layers_13_fc2_bias_to_fp16, weight = trunk_layers_13_fc2_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<fp16, [390, 896]> input_177_cast_fp16 = add(x = input_171_cast_fp16, y = linear_84_cast_fp16)[name = string("input_177_cast_fp16")];
            tensor<int32, [1]> input_179_axes_0 = const()[name = string("input_179_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_14_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_14_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146504640)))];
            tensor<fp16, [896]> trunk_layers_14_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_14_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146506496)))];
            tensor<fp16, [390, 896]> input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = trunk_layers_14_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_14_attn_ln_weight_to_fp16, x = input_177_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_14_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146508352))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147311232))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_14_q_bias_to_fp16 = const()[name = string("trunk_layers_14_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147313088)))];
            tensor<fp16, [390, 896]> linear_85_cast_fp16 = linear(bias = trunk_layers_14_q_bias_to_fp16, weight = trunk_layers_14_q_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [3]> var_911 = const()[name = string("op_911"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_912_cast_fp16 = reshape(shape = var_911, x = linear_85_cast_fp16)[name = string("op_912_cast_fp16")];
            tensor<int32, [3]> var_913_perm_0 = const()[name = string("op_913_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_29_axes_0 = const()[name = string("q_29_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_913_cast_fp16 = transpose(perm = var_913_perm_0, x = var_912_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 14, 390, 64]> q_29_cast_fp16 = expand_dims(axes = q_29_axes_0, x = var_913_cast_fp16)[name = string("q_29_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_14_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147314944))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148117824))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_14_k_bias_to_fp16 = const()[name = string("trunk_layers_14_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148119680)))];
            tensor<fp16, [390, 896]> linear_86_cast_fp16 = linear(bias = trunk_layers_14_k_bias_to_fp16, weight = trunk_layers_14_k_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [3]> var_918 = const()[name = string("op_918"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_919_cast_fp16 = reshape(shape = var_918, x = linear_86_cast_fp16)[name = string("op_919_cast_fp16")];
            tensor<int32, [3]> var_920_perm_0 = const()[name = string("op_920_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_29_axes_0 = const()[name = string("k_29_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_920_cast_fp16 = transpose(perm = var_920_perm_0, x = var_919_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 14, 390, 64]> k_29_cast_fp16 = expand_dims(axes = k_29_axes_0, x = var_920_cast_fp16)[name = string("k_29_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_14_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148121536))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148924416))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_14_v_bias_to_fp16 = const()[name = string("trunk_layers_14_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148926272)))];
            tensor<fp16, [390, 896]> linear_87_cast_fp16 = linear(bias = trunk_layers_14_v_bias_to_fp16, weight = trunk_layers_14_v_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<int32, [3]> var_925 = const()[name = string("op_925"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_926_cast_fp16 = reshape(shape = var_925, x = linear_87_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [3]> var_927_perm_0 = const()[name = string("op_927_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_29_axes_0 = const()[name = string("v_29_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_927_cast_fp16 = transpose(perm = var_927_perm_0, x = var_926_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 14, 390, 64]> v_29_cast_fp16 = expand_dims(axes = v_29_axes_0, x = var_927_cast_fp16)[name = string("v_29_cast_fp16")];
            fp16 mul_14_y_0_to_fp16 = const()[name = string("mul_14_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_14_cast_fp16 = mul(x = q_29_cast_fp16, y = mul_14_y_0_to_fp16)[name = string("mul_14_cast_fp16")];
            bool matmul_14_transpose_y_0 = const()[name = string("matmul_14_transpose_y_0"), val = bool(true)];
            bool matmul_14_transpose_x_0 = const()[name = string("matmul_14_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = mul_14_cast_fp16, y = k_29_cast_fp16)[name = string("matmul_14_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_14_cast_fp16 = add(x = matmul_14_cast_fp16, y = attn_mask_to_fp16)[name = string("add_14_cast_fp16")];
            int32 softmax_14_axis_0 = const()[name = string("softmax_14_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = add_14_cast_fp16)[name = string("softmax_14_cast_fp16")];
            bool a_29_transpose_x_0 = const()[name = string("a_29_transpose_x_0"), val = bool(false)];
            bool a_29_transpose_y_0 = const()[name = string("a_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_29_cast_fp16 = matmul(transpose_x = a_29_transpose_x_0, transpose_y = a_29_transpose_y_0, x = softmax_14_cast_fp16, y = v_29_cast_fp16)[name = string("a_29_cast_fp16")];
            tensor<int32, [1]> var_930_axes_0 = const()[name = string("op_930_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_930_cast_fp16 = squeeze(axes = var_930_axes_0, x = a_29_cast_fp16)[name = string("op_930_cast_fp16")];
            tensor<int32, [3]> var_931_perm_0 = const()[name = string("op_931_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_932 = const()[name = string("op_932"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_931_cast_fp16 = transpose(perm = var_931_perm_0, x = var_930_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [390, 896]> input_181_cast_fp16 = reshape(shape = var_932, x = var_931_cast_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_14_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148928128))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149731008))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_14_out_bias_to_fp16 = const()[name = string("trunk_layers_14_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149732864)))];
            tensor<fp16, [390, 896]> linear_88_cast_fp16 = linear(bias = trunk_layers_14_out_bias_to_fp16, weight = trunk_layers_14_out_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [390, 896]> input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_88_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<int32, [1]> input_185_axes_0 = const()[name = string("input_185_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_14_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_14_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149734720)))];
            tensor<fp16, [896]> trunk_layers_14_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_14_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149736576)))];
            tensor<fp16, [390, 896]> input_185_cast_fp16 = layer_norm(axes = input_185_axes_0, beta = trunk_layers_14_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_14_mlp_ln_weight_to_fp16, x = input_183_cast_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_14_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149738432))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152949760))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_14_fc1_bias_to_fp16 = const()[name = string("trunk_layers_14_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152956992)))];
            tensor<fp16, [390, 3584]> linear_89_cast_fp16 = linear(bias = trunk_layers_14_fc1_bias_to_fp16, weight = trunk_layers_14_fc1_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_89_cast_fp16")];
            string input_187_mode_0 = const()[name = string("input_187_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_187_cast_fp16 = gelu(mode = input_187_mode_0, x = linear_89_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_14_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152964224))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156175552))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_14_fc2_bias_to_fp16 = const()[name = string("trunk_layers_14_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156177408)))];
            tensor<fp16, [390, 896]> linear_90_cast_fp16 = linear(bias = trunk_layers_14_fc2_bias_to_fp16, weight = trunk_layers_14_fc2_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [390, 896]> input_189_cast_fp16 = add(x = input_183_cast_fp16, y = linear_90_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<int32, [1]> input_191_axes_0 = const()[name = string("input_191_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_15_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_15_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156179264)))];
            tensor<fp16, [896]> trunk_layers_15_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_15_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156181120)))];
            tensor<fp16, [390, 896]> input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = trunk_layers_15_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_15_attn_ln_weight_to_fp16, x = input_189_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_15_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156182976))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156985856))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_15_q_bias_to_fp16 = const()[name = string("trunk_layers_15_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156987712)))];
            tensor<fp16, [390, 896]> linear_91_cast_fp16 = linear(bias = trunk_layers_15_q_bias_to_fp16, weight = trunk_layers_15_q_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [3]> var_966 = const()[name = string("op_966"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_967_cast_fp16 = reshape(shape = var_966, x = linear_91_cast_fp16)[name = string("op_967_cast_fp16")];
            tensor<int32, [3]> var_968_perm_0 = const()[name = string("op_968_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_31_axes_0 = const()[name = string("q_31_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_968_cast_fp16 = transpose(perm = var_968_perm_0, x = var_967_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 14, 390, 64]> q_31_cast_fp16 = expand_dims(axes = q_31_axes_0, x = var_968_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_15_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156989568))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157792448))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_15_k_bias_to_fp16 = const()[name = string("trunk_layers_15_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157794304)))];
            tensor<fp16, [390, 896]> linear_92_cast_fp16 = linear(bias = trunk_layers_15_k_bias_to_fp16, weight = trunk_layers_15_k_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [3]> var_973 = const()[name = string("op_973"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_974_cast_fp16 = reshape(shape = var_973, x = linear_92_cast_fp16)[name = string("op_974_cast_fp16")];
            tensor<int32, [3]> var_975_perm_0 = const()[name = string("op_975_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_31_axes_0 = const()[name = string("k_31_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_975_cast_fp16 = transpose(perm = var_975_perm_0, x = var_974_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 14, 390, 64]> k_31_cast_fp16 = expand_dims(axes = k_31_axes_0, x = var_975_cast_fp16)[name = string("k_31_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_15_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157796160))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158599040))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_15_v_bias_to_fp16 = const()[name = string("trunk_layers_15_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158600896)))];
            tensor<fp16, [390, 896]> linear_93_cast_fp16 = linear(bias = trunk_layers_15_v_bias_to_fp16, weight = trunk_layers_15_v_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [3]> var_980 = const()[name = string("op_980"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_981_cast_fp16 = reshape(shape = var_980, x = linear_93_cast_fp16)[name = string("op_981_cast_fp16")];
            tensor<int32, [3]> var_982_perm_0 = const()[name = string("op_982_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_31_axes_0 = const()[name = string("v_31_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_982_cast_fp16 = transpose(perm = var_982_perm_0, x = var_981_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 14, 390, 64]> v_31_cast_fp16 = expand_dims(axes = v_31_axes_0, x = var_982_cast_fp16)[name = string("v_31_cast_fp16")];
            fp16 mul_15_y_0_to_fp16 = const()[name = string("mul_15_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_15_cast_fp16 = mul(x = q_31_cast_fp16, y = mul_15_y_0_to_fp16)[name = string("mul_15_cast_fp16")];
            bool matmul_15_transpose_y_0 = const()[name = string("matmul_15_transpose_y_0"), val = bool(true)];
            bool matmul_15_transpose_x_0 = const()[name = string("matmul_15_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = mul_15_cast_fp16, y = k_31_cast_fp16)[name = string("matmul_15_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_15_cast_fp16 = add(x = matmul_15_cast_fp16, y = attn_mask_to_fp16)[name = string("add_15_cast_fp16")];
            int32 softmax_15_axis_0 = const()[name = string("softmax_15_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = add_15_cast_fp16)[name = string("softmax_15_cast_fp16")];
            bool a_31_transpose_x_0 = const()[name = string("a_31_transpose_x_0"), val = bool(false)];
            bool a_31_transpose_y_0 = const()[name = string("a_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_31_cast_fp16 = matmul(transpose_x = a_31_transpose_x_0, transpose_y = a_31_transpose_y_0, x = softmax_15_cast_fp16, y = v_31_cast_fp16)[name = string("a_31_cast_fp16")];
            tensor<int32, [1]> var_985_axes_0 = const()[name = string("op_985_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_985_cast_fp16 = squeeze(axes = var_985_axes_0, x = a_31_cast_fp16)[name = string("op_985_cast_fp16")];
            tensor<int32, [3]> var_986_perm_0 = const()[name = string("op_986_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_987 = const()[name = string("op_987"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_986_cast_fp16 = transpose(perm = var_986_perm_0, x = var_985_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [390, 896]> input_193_cast_fp16 = reshape(shape = var_987, x = var_986_cast_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_15_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158602752))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159405632))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_15_out_bias_to_fp16 = const()[name = string("trunk_layers_15_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159407488)))];
            tensor<fp16, [390, 896]> linear_94_cast_fp16 = linear(bias = trunk_layers_15_out_bias_to_fp16, weight = trunk_layers_15_out_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [390, 896]> input_195_cast_fp16 = add(x = input_189_cast_fp16, y = linear_94_cast_fp16)[name = string("input_195_cast_fp16")];
            tensor<int32, [1]> input_197_axes_0 = const()[name = string("input_197_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_15_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_15_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159409344)))];
            tensor<fp16, [896]> trunk_layers_15_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_15_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159411200)))];
            tensor<fp16, [390, 896]> input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = trunk_layers_15_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_15_mlp_ln_weight_to_fp16, x = input_195_cast_fp16)[name = string("input_197_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_15_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159413056))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162624384))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_15_fc1_bias_to_fp16 = const()[name = string("trunk_layers_15_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162631616)))];
            tensor<fp16, [390, 3584]> linear_95_cast_fp16 = linear(bias = trunk_layers_15_fc1_bias_to_fp16, weight = trunk_layers_15_fc1_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_95_cast_fp16")];
            string input_199_mode_0 = const()[name = string("input_199_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = linear_95_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_15_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162638848))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165850176))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_15_fc2_bias_to_fp16 = const()[name = string("trunk_layers_15_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165852032)))];
            tensor<fp16, [390, 896]> linear_96_cast_fp16 = linear(bias = trunk_layers_15_fc2_bias_to_fp16, weight = trunk_layers_15_fc2_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [390, 896]> input_201_cast_fp16 = add(x = input_195_cast_fp16, y = linear_96_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<int32, [1]> input_203_axes_0 = const()[name = string("input_203_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_16_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_16_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165853888)))];
            tensor<fp16, [896]> trunk_layers_16_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_16_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165855744)))];
            tensor<fp16, [390, 896]> input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = trunk_layers_16_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_16_attn_ln_weight_to_fp16, x = input_201_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_16_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165857600))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166660480))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_16_q_bias_to_fp16 = const()[name = string("trunk_layers_16_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166662336)))];
            tensor<fp16, [390, 896]> linear_97_cast_fp16 = linear(bias = trunk_layers_16_q_bias_to_fp16, weight = trunk_layers_16_q_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<int32, [3]> var_1021 = const()[name = string("op_1021"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_1022_cast_fp16 = reshape(shape = var_1021, x = linear_97_cast_fp16)[name = string("op_1022_cast_fp16")];
            tensor<int32, [3]> var_1023_perm_0 = const()[name = string("op_1023_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_33_axes_0 = const()[name = string("q_33_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_1023_cast_fp16 = transpose(perm = var_1023_perm_0, x = var_1022_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 14, 390, 64]> q_33_cast_fp16 = expand_dims(axes = q_33_axes_0, x = var_1023_cast_fp16)[name = string("q_33_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_16_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166664192))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167467072))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_16_k_bias_to_fp16 = const()[name = string("trunk_layers_16_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167468928)))];
            tensor<fp16, [390, 896]> linear_98_cast_fp16 = linear(bias = trunk_layers_16_k_bias_to_fp16, weight = trunk_layers_16_k_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [3]> var_1028 = const()[name = string("op_1028"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_1029_cast_fp16 = reshape(shape = var_1028, x = linear_98_cast_fp16)[name = string("op_1029_cast_fp16")];
            tensor<int32, [3]> var_1030_perm_0 = const()[name = string("op_1030_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_33_axes_0 = const()[name = string("k_33_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_1030_cast_fp16 = transpose(perm = var_1030_perm_0, x = var_1029_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 14, 390, 64]> k_33_cast_fp16 = expand_dims(axes = k_33_axes_0, x = var_1030_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_16_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167470784))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168273664))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_16_v_bias_to_fp16 = const()[name = string("trunk_layers_16_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168275520)))];
            tensor<fp16, [390, 896]> linear_99_cast_fp16 = linear(bias = trunk_layers_16_v_bias_to_fp16, weight = trunk_layers_16_v_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [3]> var_1035 = const()[name = string("op_1035"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_1036_cast_fp16 = reshape(shape = var_1035, x = linear_99_cast_fp16)[name = string("op_1036_cast_fp16")];
            tensor<int32, [3]> var_1037_perm_0 = const()[name = string("op_1037_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_33_axes_0 = const()[name = string("v_33_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_1037_cast_fp16 = transpose(perm = var_1037_perm_0, x = var_1036_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 14, 390, 64]> v_33_cast_fp16 = expand_dims(axes = v_33_axes_0, x = var_1037_cast_fp16)[name = string("v_33_cast_fp16")];
            fp16 mul_16_y_0_to_fp16 = const()[name = string("mul_16_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_16_cast_fp16 = mul(x = q_33_cast_fp16, y = mul_16_y_0_to_fp16)[name = string("mul_16_cast_fp16")];
            bool matmul_16_transpose_y_0 = const()[name = string("matmul_16_transpose_y_0"), val = bool(true)];
            bool matmul_16_transpose_x_0 = const()[name = string("matmul_16_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = mul_16_cast_fp16, y = k_33_cast_fp16)[name = string("matmul_16_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_16_cast_fp16 = add(x = matmul_16_cast_fp16, y = attn_mask_to_fp16)[name = string("add_16_cast_fp16")];
            int32 softmax_16_axis_0 = const()[name = string("softmax_16_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_16_cast_fp16 = softmax(axis = softmax_16_axis_0, x = add_16_cast_fp16)[name = string("softmax_16_cast_fp16")];
            bool a_33_transpose_x_0 = const()[name = string("a_33_transpose_x_0"), val = bool(false)];
            bool a_33_transpose_y_0 = const()[name = string("a_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_33_cast_fp16 = matmul(transpose_x = a_33_transpose_x_0, transpose_y = a_33_transpose_y_0, x = softmax_16_cast_fp16, y = v_33_cast_fp16)[name = string("a_33_cast_fp16")];
            tensor<int32, [1]> var_1040_axes_0 = const()[name = string("op_1040_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_1040_cast_fp16 = squeeze(axes = var_1040_axes_0, x = a_33_cast_fp16)[name = string("op_1040_cast_fp16")];
            tensor<int32, [3]> var_1041_perm_0 = const()[name = string("op_1041_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_1042 = const()[name = string("op_1042"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_1041_cast_fp16 = transpose(perm = var_1041_perm_0, x = var_1040_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [390, 896]> input_205_cast_fp16 = reshape(shape = var_1042, x = var_1041_cast_fp16)[name = string("input_205_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_16_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168277376))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169080256))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_16_out_bias_to_fp16 = const()[name = string("trunk_layers_16_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169082112)))];
            tensor<fp16, [390, 896]> linear_100_cast_fp16 = linear(bias = trunk_layers_16_out_bias_to_fp16, weight = trunk_layers_16_out_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<fp16, [390, 896]> input_207_cast_fp16 = add(x = input_201_cast_fp16, y = linear_100_cast_fp16)[name = string("input_207_cast_fp16")];
            tensor<int32, [1]> input_209_axes_0 = const()[name = string("input_209_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_16_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_16_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169083968)))];
            tensor<fp16, [896]> trunk_layers_16_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_16_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169085824)))];
            tensor<fp16, [390, 896]> input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = trunk_layers_16_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_16_mlp_ln_weight_to_fp16, x = input_207_cast_fp16)[name = string("input_209_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_16_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169087680))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172299008))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_16_fc1_bias_to_fp16 = const()[name = string("trunk_layers_16_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172306240)))];
            tensor<fp16, [390, 3584]> linear_101_cast_fp16 = linear(bias = trunk_layers_16_fc1_bias_to_fp16, weight = trunk_layers_16_fc1_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_101_cast_fp16")];
            string input_211_mode_0 = const()[name = string("input_211_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_211_cast_fp16 = gelu(mode = input_211_mode_0, x = linear_101_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_16_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172313472))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175524800))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_16_fc2_bias_to_fp16 = const()[name = string("trunk_layers_16_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175526656)))];
            tensor<fp16, [390, 896]> linear_102_cast_fp16 = linear(bias = trunk_layers_16_fc2_bias_to_fp16, weight = trunk_layers_16_fc2_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [390, 896]> input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_102_cast_fp16)[name = string("input_213_cast_fp16")];
            tensor<int32, [1]> input_215_axes_0 = const()[name = string("input_215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_17_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_17_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175528512)))];
            tensor<fp16, [896]> trunk_layers_17_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_17_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175530368)))];
            tensor<fp16, [390, 896]> input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = trunk_layers_17_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_17_attn_ln_weight_to_fp16, x = input_213_cast_fp16)[name = string("input_215_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_17_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175532224))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176335104))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_17_q_bias_to_fp16 = const()[name = string("trunk_layers_17_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176336960)))];
            tensor<fp16, [390, 896]> linear_103_cast_fp16 = linear(bias = trunk_layers_17_q_bias_to_fp16, weight = trunk_layers_17_q_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<int32, [3]> var_1076 = const()[name = string("op_1076"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_1077_cast_fp16 = reshape(shape = var_1076, x = linear_103_cast_fp16)[name = string("op_1077_cast_fp16")];
            tensor<int32, [3]> var_1078_perm_0 = const()[name = string("op_1078_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_35_axes_0 = const()[name = string("q_35_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_1078_cast_fp16 = transpose(perm = var_1078_perm_0, x = var_1077_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 14, 390, 64]> q_35_cast_fp16 = expand_dims(axes = q_35_axes_0, x = var_1078_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_17_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176338816))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177141696))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_17_k_bias_to_fp16 = const()[name = string("trunk_layers_17_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177143552)))];
            tensor<fp16, [390, 896]> linear_104_cast_fp16 = linear(bias = trunk_layers_17_k_bias_to_fp16, weight = trunk_layers_17_k_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<int32, [3]> var_1083 = const()[name = string("op_1083"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_1084_cast_fp16 = reshape(shape = var_1083, x = linear_104_cast_fp16)[name = string("op_1084_cast_fp16")];
            tensor<int32, [3]> var_1085_perm_0 = const()[name = string("op_1085_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_35_axes_0 = const()[name = string("k_35_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_1085_cast_fp16 = transpose(perm = var_1085_perm_0, x = var_1084_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 14, 390, 64]> k_35_cast_fp16 = expand_dims(axes = k_35_axes_0, x = var_1085_cast_fp16)[name = string("k_35_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_17_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177145408))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177948288))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_17_v_bias_to_fp16 = const()[name = string("trunk_layers_17_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177950144)))];
            tensor<fp16, [390, 896]> linear_105_cast_fp16 = linear(bias = trunk_layers_17_v_bias_to_fp16, weight = trunk_layers_17_v_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [3]> var_1090 = const()[name = string("op_1090"), val = tensor<int32, [3]>([390, 14, 64])];
            tensor<fp16, [390, 14, 64]> var_1091_cast_fp16 = reshape(shape = var_1090, x = linear_105_cast_fp16)[name = string("op_1091_cast_fp16")];
            tensor<int32, [3]> var_1092_perm_0 = const()[name = string("op_1092_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_35_axes_0 = const()[name = string("v_35_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_1092_cast_fp16 = transpose(perm = var_1092_perm_0, x = var_1091_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 14, 390, 64]> v_35_cast_fp16 = expand_dims(axes = v_35_axes_0, x = var_1092_cast_fp16)[name = string("v_35_cast_fp16")];
            fp16 mul_17_y_0_to_fp16 = const()[name = string("mul_17_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 390, 64]> mul_17_cast_fp16 = mul(x = q_35_cast_fp16, y = mul_17_y_0_to_fp16)[name = string("mul_17_cast_fp16")];
            bool matmul_17_transpose_y_0 = const()[name = string("matmul_17_transpose_y_0"), val = bool(true)];
            bool matmul_17_transpose_x_0 = const()[name = string("matmul_17_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 390]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = mul_17_cast_fp16, y = k_35_cast_fp16)[name = string("matmul_17_cast_fp16")];
            tensor<fp16, [1, 14, 390, 390]> add_17_cast_fp16 = add(x = matmul_17_cast_fp16, y = attn_mask_to_fp16)[name = string("add_17_cast_fp16")];
            int32 softmax_17_axis_0 = const()[name = string("softmax_17_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 390, 390]> softmax_17_cast_fp16 = softmax(axis = softmax_17_axis_0, x = add_17_cast_fp16)[name = string("softmax_17_cast_fp16")];
            bool a_transpose_x_0 = const()[name = string("a_transpose_x_0"), val = bool(false)];
            bool a_transpose_y_0 = const()[name = string("a_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 390, 64]> a_cast_fp16 = matmul(transpose_x = a_transpose_x_0, transpose_y = a_transpose_y_0, x = softmax_17_cast_fp16, y = v_35_cast_fp16)[name = string("a_cast_fp16")];
            tensor<int32, [1]> var_1095_axes_0 = const()[name = string("op_1095_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 390, 64]> var_1095_cast_fp16 = squeeze(axes = var_1095_axes_0, x = a_cast_fp16)[name = string("op_1095_cast_fp16")];
            tensor<int32, [3]> var_1096_perm_0 = const()[name = string("op_1096_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_1097 = const()[name = string("op_1097"), val = tensor<int32, [2]>([390, 896])];
            tensor<fp16, [390, 14, 64]> var_1096_cast_fp16 = transpose(perm = var_1096_perm_0, x = var_1095_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [390, 896]> input_217_cast_fp16 = reshape(shape = var_1097, x = var_1096_cast_fp16)[name = string("input_217_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_17_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177952000))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178754880))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_17_out_bias_to_fp16 = const()[name = string("trunk_layers_17_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178756736)))];
            tensor<fp16, [390, 896]> linear_106_cast_fp16 = linear(bias = trunk_layers_17_out_bias_to_fp16, weight = trunk_layers_17_out_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<fp16, [390, 896]> input_219_cast_fp16 = add(x = input_213_cast_fp16, y = linear_106_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<int32, [1]> input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_17_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_17_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178758592)))];
            tensor<fp16, [896]> trunk_layers_17_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_17_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178760448)))];
            tensor<fp16, [390, 896]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = trunk_layers_17_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_17_mlp_ln_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_17_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178762304))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181973632))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_17_fc1_bias_to_fp16 = const()[name = string("trunk_layers_17_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181980864)))];
            tensor<fp16, [390, 3584]> linear_107_cast_fp16 = linear(bias = trunk_layers_17_fc1_bias_to_fp16, weight = trunk_layers_17_fc1_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_107_cast_fp16")];
            string input_223_mode_0 = const()[name = string("input_223_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 3584]> input_223_cast_fp16 = gelu(mode = input_223_mode_0, x = linear_107_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_17_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181988096))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185199424))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_17_fc2_bias_to_fp16 = const()[name = string("trunk_layers_17_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185201280)))];
            tensor<fp16, [390, 896]> linear_108_cast_fp16 = linear(bias = trunk_layers_17_fc2_bias_to_fp16, weight = trunk_layers_17_fc2_weight_to_fp16_quantized, x = input_223_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [390, 896]> input_225_cast_fp16 = add(x = input_219_cast_fp16, y = linear_108_cast_fp16)[name = string("input_225_cast_fp16")];
            tensor<int32, [1]> input_227_axes_0 = const()[name = string("input_227_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_ln_post_weight_to_fp16 = const()[name = string("trunk_ln_post_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185203136)))];
            tensor<fp16, [896]> trunk_ln_post_bias_to_fp16 = const()[name = string("trunk_ln_post_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185204992)))];
            tensor<fp16, [390, 896]> input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = trunk_ln_post_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_ln_post_weight_to_fp16, x = input_225_cast_fp16)[name = string("input_227_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_proj1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_proj1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185206848))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186009728))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_proj1_bias_to_fp16 = const()[name = string("trunk_proj1_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186011584)))];
            tensor<fp16, [390, 896]> linear_109_cast_fp16 = linear(bias = trunk_proj1_bias_to_fp16, weight = trunk_proj1_weight_to_fp16_quantized, x = input_227_cast_fp16)[name = string("linear_109_cast_fp16")];
            string input_229_mode_0 = const()[name = string("input_229_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 896]> input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = linear_109_cast_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [1024, 896]> trunk_proj2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_proj2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186013440))), scale = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186932096))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186931008)))];
            tensor<fp16, [1024]> trunk_proj2_bias_to_fp16 = const()[name = string("trunk_proj2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186934208)))];
            tensor<fp16, [390, 1024]> linear_110_cast_fp16 = linear(bias = trunk_proj2_bias_to_fp16, weight = trunk_proj2_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<int32, [1]> input_233_axes_0 = const()[name = string("input_233_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> trunk_tower_0_norm_weight_to_fp16 = const()[name = string("trunk_tower_0_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186936320)))];
            tensor<fp16, [1024]> trunk_tower_0_norm_bias_to_fp16 = const()[name = string("trunk_tower_0_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186938432)))];
            tensor<fp16, [390, 1024]> input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = trunk_tower_0_norm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_tower_0_norm_weight_to_fp16, x = linear_110_cast_fp16)[name = string("input_233_cast_fp16")];
            tensor<fp16, [4096, 1024]> trunk_tower_0_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_0_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186940544))), scale = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191139072))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191134912)))];
            tensor<fp16, [4096]> trunk_tower_0_fc1_bias_to_fp16 = const()[name = string("trunk_tower_0_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191147328)))];
            tensor<fp16, [390, 4096]> linear_111_cast_fp16 = linear(bias = trunk_tower_0_fc1_bias_to_fp16, weight = trunk_tower_0_fc1_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_111_cast_fp16")];
            string input_235_mode_0 = const()[name = string("input_235_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 4096]> input_235_cast_fp16 = gelu(mode = input_235_mode_0, x = linear_111_cast_fp16)[name = string("input_235_cast_fp16")];
            tensor<fp16, [1024, 4096]> trunk_tower_0_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_0_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191155584))), scale = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195349952))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186931008)))];
            tensor<fp16, [1024]> trunk_tower_0_fc2_bias_to_fp16 = const()[name = string("trunk_tower_0_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195352064)))];
            tensor<fp16, [390, 1024]> linear_112_cast_fp16 = linear(bias = trunk_tower_0_fc2_bias_to_fp16, weight = trunk_tower_0_fc2_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<fp16, [390, 1024]> input_237_cast_fp16 = add(x = linear_110_cast_fp16, y = linear_112_cast_fp16)[name = string("input_237_cast_fp16")];
            tensor<int32, [1]> input_239_axes_0 = const()[name = string("input_239_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> trunk_tower_1_norm_weight_to_fp16 = const()[name = string("trunk_tower_1_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195354176)))];
            tensor<fp16, [1024]> trunk_tower_1_norm_bias_to_fp16 = const()[name = string("trunk_tower_1_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195356288)))];
            tensor<fp16, [390, 1024]> input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = trunk_tower_1_norm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_tower_1_norm_weight_to_fp16, x = input_237_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [4096, 1024]> trunk_tower_1_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_1_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195358400))), scale = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199552768))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191134912)))];
            tensor<fp16, [4096]> trunk_tower_1_fc1_bias_to_fp16 = const()[name = string("trunk_tower_1_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199561024)))];
            tensor<fp16, [390, 4096]> linear_113_cast_fp16 = linear(bias = trunk_tower_1_fc1_bias_to_fp16, weight = trunk_tower_1_fc1_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_113_cast_fp16")];
            string input_241_mode_0 = const()[name = string("input_241_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 4096]> input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_113_cast_fp16)[name = string("input_241_cast_fp16")];
            tensor<fp16, [1024, 4096]> trunk_tower_1_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_1_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199569280))), scale = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203763648))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186931008)))];
            tensor<fp16, [1024]> trunk_tower_1_fc2_bias_to_fp16 = const()[name = string("trunk_tower_1_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203765760)))];
            tensor<fp16, [390, 1024]> linear_114_cast_fp16 = linear(bias = trunk_tower_1_fc2_bias_to_fp16, weight = trunk_tower_1_fc2_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<fp16, [390, 1024]> input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_114_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<int32, [1]> input_245_axes_0 = const()[name = string("input_245_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> trunk_tower_2_norm_weight_to_fp16 = const()[name = string("trunk_tower_2_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203767872)))];
            tensor<fp16, [1024]> trunk_tower_2_norm_bias_to_fp16 = const()[name = string("trunk_tower_2_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203769984)))];
            tensor<fp16, [390, 1024]> input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = trunk_tower_2_norm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_tower_2_norm_weight_to_fp16, x = input_243_cast_fp16)[name = string("input_245_cast_fp16")];
            tensor<fp16, [4096, 1024]> trunk_tower_2_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_2_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203772096))), scale = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207966464))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191134912)))];
            tensor<fp16, [4096]> trunk_tower_2_fc1_bias_to_fp16 = const()[name = string("trunk_tower_2_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207974720)))];
            tensor<fp16, [390, 4096]> linear_115_cast_fp16 = linear(bias = trunk_tower_2_fc1_bias_to_fp16, weight = trunk_tower_2_fc1_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_115_cast_fp16")];
            string input_247_mode_0 = const()[name = string("input_247_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 4096]> input_247_cast_fp16 = gelu(mode = input_247_mode_0, x = linear_115_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<fp16, [1024, 4096]> trunk_tower_2_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_2_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207982976))), scale = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212177344))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186931008)))];
            tensor<fp16, [1024]> trunk_tower_2_fc2_bias_to_fp16 = const()[name = string("trunk_tower_2_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212179456)))];
            tensor<fp16, [390, 1024]> linear_116_cast_fp16 = linear(bias = trunk_tower_2_fc2_bias_to_fp16, weight = trunk_tower_2_fc2_weight_to_fp16_quantized, x = input_247_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [390, 1024]> input_249_cast_fp16 = add(x = input_243_cast_fp16, y = linear_116_cast_fp16)[name = string("input_249_cast_fp16")];
            tensor<int32, [1]> input_251_axes_0 = const()[name = string("input_251_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> trunk_tower_3_norm_weight_to_fp16 = const()[name = string("trunk_tower_3_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212181568)))];
            tensor<fp16, [1024]> trunk_tower_3_norm_bias_to_fp16 = const()[name = string("trunk_tower_3_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212183680)))];
            tensor<fp16, [390, 1024]> input_251_cast_fp16 = layer_norm(axes = input_251_axes_0, beta = trunk_tower_3_norm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_tower_3_norm_weight_to_fp16, x = input_249_cast_fp16)[name = string("input_251_cast_fp16")];
            tensor<fp16, [4096, 1024]> trunk_tower_3_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_3_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212185792))), scale = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216380160))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191134912)))];
            tensor<fp16, [4096]> trunk_tower_3_fc1_bias_to_fp16 = const()[name = string("trunk_tower_3_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216388416)))];
            tensor<fp16, [390, 4096]> linear_117_cast_fp16 = linear(bias = trunk_tower_3_fc1_bias_to_fp16, weight = trunk_tower_3_fc1_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_117_cast_fp16")];
            string input_253_mode_0 = const()[name = string("input_253_mode_0"), val = string("EXACT")];
            tensor<fp16, [390, 4096]> input_253_cast_fp16 = gelu(mode = input_253_mode_0, x = linear_117_cast_fp16)[name = string("input_253_cast_fp16")];
            tensor<fp16, [1024, 4096]> trunk_tower_3_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_3_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216396672))), scale = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220591040))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186931008)))];
            tensor<fp16, [1024]> trunk_tower_3_fc2_bias_to_fp16 = const()[name = string("trunk_tower_3_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220593152)))];
            tensor<fp16, [390, 1024]> linear_118_cast_fp16 = linear(bias = trunk_tower_3_fc2_bias_to_fp16, weight = trunk_tower_3_fc2_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [390, 1024]> input_cast_fp16 = add(x = input_249_cast_fp16, y = linear_118_cast_fp16)[name = string("input_cast_fp16")];
            tensor<int32, [1]> var_1189_axes_0 = const()[name = string("op_1189_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> trunk_final_norm_weight_to_fp16 = const()[name = string("trunk_final_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220595264)))];
            tensor<fp16, [1024]> trunk_final_norm_bias_to_fp16 = const()[name = string("trunk_final_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220597376)))];
            tensor<fp16, [390, 1024]> hidden = layer_norm(axes = var_1189_axes_0, beta = trunk_final_norm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_final_norm_weight_to_fp16, x = input_cast_fp16)[name = string("op_1189_cast_fp16")];
        } -> (hidden);
    func tower_5s<ios18>(tensor<fp32, [1, 1, 65, 65]> attn_mask, tensor<fp32, [1, 128, 500]> audios) {
            tensor<int32, [3]> var_25_begin_0 = const()[name = string("op_25_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_25_end_0 = const()[name = string("op_25_end_0"), val = tensor<int32, [3]>([1, 128, 500])];
            tensor<bool, [3]> var_25_end_mask_0 = const()[name = string("op_25_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> var_25_squeeze_mask_0 = const()[name = string("op_25_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            string audios_to_fp16_dtype_0 = const()[name = string("audios_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 500]> audios_to_fp16 = cast(dtype = audios_to_fp16_dtype_0, x = audios)[name = string("cast_1")];
            tensor<fp16, [128, 500]> var_25_cast_fp16 = slice_by_index(begin = var_25_begin_0, end = var_25_end_0, end_mask = var_25_end_mask_0, squeeze_mask = var_25_squeeze_mask_0, x = audios_to_fp16)[name = string("op_25_cast_fp16")];
            tensor<int32, [2]> x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor<int32, [2]>([1, 0])];
            tensor<int32, [3]> var_27 = const()[name = string("op_27"), val = tensor<int32, [3]>([5, 100, 128])];
            tensor<fp16, [500, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = var_25_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [5, 100, 128]> x_3_cast_fp16 = reshape(shape = var_27, x = x_1_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<int32, [3]> var_29 = const()[name = string("op_29"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_1_axes_0 = const()[name = string("input_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [5, 128, 100]> var_30_cast_fp16 = transpose(perm = var_29, x = x_3_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [5, 1, 128, 100]> input_1_cast_fp16 = expand_dims(axes = input_1_axes_0, x = var_30_cast_fp16)[name = string("input_1_cast_fp16")];
            string var_38_pad_type_0 = const()[name = string("op_38_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_38_pad_0 = const()[name = string("op_38_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> var_38_strides_0 = const()[name = string("op_38_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_38_dilations_0 = const()[name = string("op_38_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_38_groups_0 = const()[name = string("op_38_groups_0"), val = int32(1)];
            tensor<fp16, [480, 1, 3, 3]> frontend_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("frontend_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [480, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5056))), zero_point = tensor<int8, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4480)))];
            tensor<fp16, [480]> frontend_conv1_bias_to_fp16 = const()[name = string("frontend_conv1_bias_to_fp16"), val = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6080)))];
            tensor<fp16, [5, 480, 64, 50]> var_38_cast_fp16 = conv(bias = frontend_conv1_bias_to_fp16, dilations = var_38_dilations_0, groups = var_38_groups_0, pad = var_38_pad_0, pad_type = var_38_pad_type_0, strides = var_38_strides_0, weight = frontend_conv1_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("op_38_cast_fp16")];
            string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("EXACT")];
            tensor<fp16, [5, 480, 64, 50]> input_3_cast_fp16 = gelu(mode = input_3_mode_0, x = var_38_cast_fp16)[name = string("input_3_cast_fp16")];
            string var_46_pad_type_0 = const()[name = string("op_46_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_46_pad_0 = const()[name = string("op_46_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> var_46_strides_0 = const()[name = string("op_46_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_46_dilations_0 = const()[name = string("op_46_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_46_groups_0 = const()[name = string("op_46_groups_0"), val = int32(1)];
            tensor<fp16, [480, 480, 3, 3]> frontend_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("frontend_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [480, 480, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7104))), scale = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2080768))), zero_point = tensor<int8, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4480)))];
            tensor<fp16, [480]> frontend_conv2_bias_to_fp16 = const()[name = string("frontend_conv2_bias_to_fp16"), val = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2081792)))];
            tensor<fp16, [5, 480, 32, 25]> var_46_cast_fp16 = conv(bias = frontend_conv2_bias_to_fp16, dilations = var_46_dilations_0, groups = var_46_groups_0, pad = var_46_pad_0, pad_type = var_46_pad_type_0, strides = var_46_strides_0, weight = frontend_conv2_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("op_46_cast_fp16")];
            string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("EXACT")];
            tensor<fp16, [5, 480, 32, 25]> input_5_cast_fp16 = gelu(mode = input_5_mode_0, x = var_46_cast_fp16)[name = string("input_5_cast_fp16")];
            string var_54_pad_type_0 = const()[name = string("op_54_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_54_pad_0 = const()[name = string("op_54_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> var_54_strides_0 = const()[name = string("op_54_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> var_54_dilations_0 = const()[name = string("op_54_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_54_groups_0 = const()[name = string("op_54_groups_0"), val = int32(1)];
            tensor<fp16, [480, 480, 3, 3]> frontend_conv3_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("frontend_conv3_weight_to_fp16_quantized"), quantized_data = tensor<int8, [480, 480, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2082816))), scale = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4156480))), zero_point = tensor<int8, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4480)))];
            tensor<fp16, [480]> frontend_conv3_bias_to_fp16 = const()[name = string("frontend_conv3_bias_to_fp16"), val = tensor<fp16, [480]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4157504)))];
            tensor<fp16, [5, 480, 16, 13]> var_54_cast_fp16 = conv(bias = frontend_conv3_bias_to_fp16, dilations = var_54_dilations_0, groups = var_54_groups_0, pad = var_54_pad_0, pad_type = var_54_pad_type_0, strides = var_54_strides_0, weight = frontend_conv3_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("op_54_cast_fp16")];
            string x_5_mode_0 = const()[name = string("x_5_mode_0"), val = string("EXACT")];
            tensor<fp16, [5, 480, 16, 13]> x_5_cast_fp16 = gelu(mode = x_5_mode_0, x = var_54_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<int32, [4]> var_56 = const()[name = string("op_56"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [3]> var_58 = const()[name = string("op_58"), val = tensor<int32, [3]>([5, 13, 7680])];
            tensor<fp16, [5, 13, 480, 16]> var_57_cast_fp16 = transpose(perm = var_56, x = x_5_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [5, 13, 7680]> input_7_cast_fp16 = reshape(shape = var_58, x = var_57_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [896, 7680]> frontend_conv_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("frontend_conv_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 7680]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4158528))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11040832))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11042688)))];
            tensor<fp16, [5, 13, 896]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = frontend_conv_out_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<fp16, [1, 13, 896]> frontend_pos_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("frontend_pos_to_fp16_quantized"), quantized_data = tensor<int8, [1, 13, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11044544))), scale = fp16(0x1.02p-7), zero_point = int8(0)];
            tensor<fp16, [5, 13, 896]> x_cast_fp16 = add(x = linear_0_cast_fp16, y = frontend_pos_to_fp16_quantized)[name = string("x_cast_fp16")];
            tensor<int32, [2]> var_63 = const()[name = string("op_63"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 896]> input_9_cast_fp16 = reshape(shape = var_63, x = x_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<int32, [1]> input_11_axes_0 = const()[name = string("input_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_0_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_0_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11056256)))];
            tensor<fp16, [896]> trunk_layers_0_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_0_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11058112)))];
            fp16 var_67_to_fp16 = const()[name = string("op_67_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [65, 896]> input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = trunk_layers_0_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_0_attn_ln_weight_to_fp16, x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_0_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11059968))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11862848))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_0_q_bias_to_fp16 = const()[name = string("trunk_layers_0_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11864704)))];
            tensor<fp16, [65, 896]> linear_1_cast_fp16 = linear(bias = trunk_layers_0_q_bias_to_fp16, weight = trunk_layers_0_q_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [3]> var_141 = const()[name = string("op_141"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_142_cast_fp16 = reshape(shape = var_141, x = linear_1_cast_fp16)[name = string("op_142_cast_fp16")];
            tensor<int32, [3]> var_143_perm_0 = const()[name = string("op_143_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_1_axes_0 = const()[name = string("q_1_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_143_cast_fp16 = transpose(perm = var_143_perm_0, x = var_142_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 14, 65, 64]> q_1_cast_fp16 = expand_dims(axes = q_1_axes_0, x = var_143_cast_fp16)[name = string("q_1_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_0_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11866560))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12669440))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_0_k_bias_to_fp16 = const()[name = string("trunk_layers_0_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12671296)))];
            tensor<fp16, [65, 896]> linear_2_cast_fp16 = linear(bias = trunk_layers_0_k_bias_to_fp16, weight = trunk_layers_0_k_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [3]> var_148 = const()[name = string("op_148"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_149_cast_fp16 = reshape(shape = var_148, x = linear_2_cast_fp16)[name = string("op_149_cast_fp16")];
            tensor<int32, [3]> var_150_perm_0 = const()[name = string("op_150_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_1_axes_0 = const()[name = string("k_1_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_150_cast_fp16 = transpose(perm = var_150_perm_0, x = var_149_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 14, 65, 64]> k_1_cast_fp16 = expand_dims(axes = k_1_axes_0, x = var_150_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_0_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12673152))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13476032))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_0_v_bias_to_fp16 = const()[name = string("trunk_layers_0_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13477888)))];
            tensor<fp16, [65, 896]> linear_3_cast_fp16 = linear(bias = trunk_layers_0_v_bias_to_fp16, weight = trunk_layers_0_v_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<int32, [3]> var_155 = const()[name = string("op_155"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_156_cast_fp16 = reshape(shape = var_155, x = linear_3_cast_fp16)[name = string("op_156_cast_fp16")];
            tensor<int32, [3]> var_157_perm_0 = const()[name = string("op_157_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_1_axes_0 = const()[name = string("v_1_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_157_cast_fp16 = transpose(perm = var_157_perm_0, x = var_156_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 14, 65, 64]> v_1_cast_fp16 = expand_dims(axes = v_1_axes_0, x = var_157_cast_fp16)[name = string("v_1_cast_fp16")];
            fp16 mul_0_y_0_to_fp16 = const()[name = string("mul_0_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_0_cast_fp16 = mul(x = q_1_cast_fp16, y = mul_0_y_0_to_fp16)[name = string("mul_0_cast_fp16")];
            bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(true)];
            bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_0_cast_fp16, y = k_1_cast_fp16)[name = string("matmul_0_cast_fp16")];
            string attn_mask_to_fp16_dtype_0 = const()[name = string("attn_mask_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1, 65, 65]> attn_mask_to_fp16 = cast(dtype = attn_mask_to_fp16_dtype_0, x = attn_mask)[name = string("cast_0")];
            tensor<fp16, [1, 14, 65, 65]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = attn_mask_to_fp16)[name = string("add_0_cast_fp16")];
            int32 softmax_0_axis_0 = const()[name = string("softmax_0_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = string("softmax_0_cast_fp16")];
            bool a_1_transpose_x_0 = const()[name = string("a_1_transpose_x_0"), val = bool(false)];
            bool a_1_transpose_y_0 = const()[name = string("a_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_1_cast_fp16 = matmul(transpose_x = a_1_transpose_x_0, transpose_y = a_1_transpose_y_0, x = softmax_0_cast_fp16, y = v_1_cast_fp16)[name = string("a_1_cast_fp16")];
            tensor<int32, [1]> var_160_axes_0 = const()[name = string("op_160_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_160_cast_fp16 = squeeze(axes = var_160_axes_0, x = a_1_cast_fp16)[name = string("op_160_cast_fp16")];
            tensor<int32, [3]> var_161_perm_0 = const()[name = string("op_161_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_162 = const()[name = string("op_162"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_161_cast_fp16 = transpose(perm = var_161_perm_0, x = var_160_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [65, 896]> input_13_cast_fp16 = reshape(shape = var_162, x = var_161_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_0_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13479744))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14282624))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_0_out_bias_to_fp16 = const()[name = string("trunk_layers_0_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14284480)))];
            tensor<fp16, [65, 896]> linear_4_cast_fp16 = linear(bias = trunk_layers_0_out_bias_to_fp16, weight = trunk_layers_0_out_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [65, 896]> input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<int32, [1]> input_17_axes_0 = const()[name = string("input_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_0_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_0_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14286336)))];
            tensor<fp16, [896]> trunk_layers_0_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_0_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14288192)))];
            tensor<fp16, [65, 896]> input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = trunk_layers_0_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_0_mlp_ln_weight_to_fp16, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_0_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14290048))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17505024))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_0_fc1_bias_to_fp16 = const()[name = string("trunk_layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17512256)))];
            tensor<fp16, [65, 3584]> linear_5_cast_fp16 = linear(bias = trunk_layers_0_fc1_bias_to_fp16, weight = trunk_layers_0_fc1_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_5_cast_fp16")];
            string input_19_mode_0 = const()[name = string("input_19_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = linear_5_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_0_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_0_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17519488))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20730816))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_0_fc2_bias_to_fp16 = const()[name = string("trunk_layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20732672)))];
            tensor<fp16, [65, 896]> linear_6_cast_fp16 = linear(bias = trunk_layers_0_fc2_bias_to_fp16, weight = trunk_layers_0_fc2_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [65, 896]> input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<int32, [1]> input_23_axes_0 = const()[name = string("input_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_1_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_1_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20734528)))];
            tensor<fp16, [896]> trunk_layers_1_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_1_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20736384)))];
            tensor<fp16, [65, 896]> input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = trunk_layers_1_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_1_attn_ln_weight_to_fp16, x = input_21_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_1_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20738240))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21541120))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_1_q_bias_to_fp16 = const()[name = string("trunk_layers_1_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21542976)))];
            tensor<fp16, [65, 896]> linear_7_cast_fp16 = linear(bias = trunk_layers_1_q_bias_to_fp16, weight = trunk_layers_1_q_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [3]> var_196 = const()[name = string("op_196"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_197_cast_fp16 = reshape(shape = var_196, x = linear_7_cast_fp16)[name = string("op_197_cast_fp16")];
            tensor<int32, [3]> var_198_perm_0 = const()[name = string("op_198_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_3_axes_0 = const()[name = string("q_3_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_198_cast_fp16 = transpose(perm = var_198_perm_0, x = var_197_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 14, 65, 64]> q_3_cast_fp16 = expand_dims(axes = q_3_axes_0, x = var_198_cast_fp16)[name = string("q_3_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_1_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21544832))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22347712))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_1_k_bias_to_fp16 = const()[name = string("trunk_layers_1_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22349568)))];
            tensor<fp16, [65, 896]> linear_8_cast_fp16 = linear(bias = trunk_layers_1_k_bias_to_fp16, weight = trunk_layers_1_k_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [3]> var_203 = const()[name = string("op_203"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_204_cast_fp16 = reshape(shape = var_203, x = linear_8_cast_fp16)[name = string("op_204_cast_fp16")];
            tensor<int32, [3]> var_205_perm_0 = const()[name = string("op_205_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_3_axes_0 = const()[name = string("k_3_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_205_cast_fp16 = transpose(perm = var_205_perm_0, x = var_204_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 14, 65, 64]> k_3_cast_fp16 = expand_dims(axes = k_3_axes_0, x = var_205_cast_fp16)[name = string("k_3_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_1_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22351424))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23154304))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_1_v_bias_to_fp16 = const()[name = string("trunk_layers_1_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23156160)))];
            tensor<fp16, [65, 896]> linear_9_cast_fp16 = linear(bias = trunk_layers_1_v_bias_to_fp16, weight = trunk_layers_1_v_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [3]> var_210 = const()[name = string("op_210"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_211_cast_fp16 = reshape(shape = var_210, x = linear_9_cast_fp16)[name = string("op_211_cast_fp16")];
            tensor<int32, [3]> var_212_perm_0 = const()[name = string("op_212_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_3_axes_0 = const()[name = string("v_3_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_212_cast_fp16 = transpose(perm = var_212_perm_0, x = var_211_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 14, 65, 64]> v_3_cast_fp16 = expand_dims(axes = v_3_axes_0, x = var_212_cast_fp16)[name = string("v_3_cast_fp16")];
            fp16 mul_1_y_0_to_fp16 = const()[name = string("mul_1_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_1_cast_fp16 = mul(x = q_3_cast_fp16, y = mul_1_y_0_to_fp16)[name = string("mul_1_cast_fp16")];
            bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(true)];
            bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_1_cast_fp16, y = k_3_cast_fp16)[name = string("matmul_1_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = attn_mask_to_fp16)[name = string("add_1_cast_fp16")];
            int32 softmax_1_axis_0 = const()[name = string("softmax_1_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = string("softmax_1_cast_fp16")];
            bool a_3_transpose_x_0 = const()[name = string("a_3_transpose_x_0"), val = bool(false)];
            bool a_3_transpose_y_0 = const()[name = string("a_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_3_cast_fp16 = matmul(transpose_x = a_3_transpose_x_0, transpose_y = a_3_transpose_y_0, x = softmax_1_cast_fp16, y = v_3_cast_fp16)[name = string("a_3_cast_fp16")];
            tensor<int32, [1]> var_215_axes_0 = const()[name = string("op_215_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_215_cast_fp16 = squeeze(axes = var_215_axes_0, x = a_3_cast_fp16)[name = string("op_215_cast_fp16")];
            tensor<int32, [3]> var_216_perm_0 = const()[name = string("op_216_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_217 = const()[name = string("op_217"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_216_cast_fp16 = transpose(perm = var_216_perm_0, x = var_215_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [65, 896]> input_25_cast_fp16 = reshape(shape = var_217, x = var_216_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_1_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23158016))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23960896))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_1_out_bias_to_fp16 = const()[name = string("trunk_layers_1_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23962752)))];
            tensor<fp16, [65, 896]> linear_10_cast_fp16 = linear(bias = trunk_layers_1_out_bias_to_fp16, weight = trunk_layers_1_out_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [65, 896]> input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_10_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<int32, [1]> input_29_axes_0 = const()[name = string("input_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_1_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_1_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23964608)))];
            tensor<fp16, [896]> trunk_layers_1_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_1_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23966464)))];
            tensor<fp16, [65, 896]> input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = trunk_layers_1_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_1_mlp_ln_weight_to_fp16, x = input_27_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_1_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23968320))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27179648))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_1_fc1_bias_to_fp16 = const()[name = string("trunk_layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27186880)))];
            tensor<fp16, [65, 3584]> linear_11_cast_fp16 = linear(bias = trunk_layers_1_fc1_bias_to_fp16, weight = trunk_layers_1_fc1_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_11_cast_fp16")];
            string input_31_mode_0 = const()[name = string("input_31_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = linear_11_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_1_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_1_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27194112))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30405440))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_1_fc2_bias_to_fp16 = const()[name = string("trunk_layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30407296)))];
            tensor<fp16, [65, 896]> linear_12_cast_fp16 = linear(bias = trunk_layers_1_fc2_bias_to_fp16, weight = trunk_layers_1_fc2_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [65, 896]> input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_12_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<int32, [1]> input_35_axes_0 = const()[name = string("input_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_2_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_2_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30409152)))];
            tensor<fp16, [896]> trunk_layers_2_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_2_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30411008)))];
            tensor<fp16, [65, 896]> input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = trunk_layers_2_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_2_attn_ln_weight_to_fp16, x = input_33_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_2_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30412864))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31215744))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_2_q_bias_to_fp16 = const()[name = string("trunk_layers_2_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31217600)))];
            tensor<fp16, [65, 896]> linear_13_cast_fp16 = linear(bias = trunk_layers_2_q_bias_to_fp16, weight = trunk_layers_2_q_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<int32, [3]> var_251 = const()[name = string("op_251"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_252_cast_fp16 = reshape(shape = var_251, x = linear_13_cast_fp16)[name = string("op_252_cast_fp16")];
            tensor<int32, [3]> var_253_perm_0 = const()[name = string("op_253_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_5_axes_0 = const()[name = string("q_5_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_253_cast_fp16 = transpose(perm = var_253_perm_0, x = var_252_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 14, 65, 64]> q_5_cast_fp16 = expand_dims(axes = q_5_axes_0, x = var_253_cast_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_2_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31219456))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32022336))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_2_k_bias_to_fp16 = const()[name = string("trunk_layers_2_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32024192)))];
            tensor<fp16, [65, 896]> linear_14_cast_fp16 = linear(bias = trunk_layers_2_k_bias_to_fp16, weight = trunk_layers_2_k_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [3]> var_258 = const()[name = string("op_258"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_259_cast_fp16 = reshape(shape = var_258, x = linear_14_cast_fp16)[name = string("op_259_cast_fp16")];
            tensor<int32, [3]> var_260_perm_0 = const()[name = string("op_260_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_5_axes_0 = const()[name = string("k_5_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_260_cast_fp16 = transpose(perm = var_260_perm_0, x = var_259_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 14, 65, 64]> k_5_cast_fp16 = expand_dims(axes = k_5_axes_0, x = var_260_cast_fp16)[name = string("k_5_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_2_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32026048))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32828928))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_2_v_bias_to_fp16 = const()[name = string("trunk_layers_2_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32830784)))];
            tensor<fp16, [65, 896]> linear_15_cast_fp16 = linear(bias = trunk_layers_2_v_bias_to_fp16, weight = trunk_layers_2_v_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [3]> var_265 = const()[name = string("op_265"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_266_cast_fp16 = reshape(shape = var_265, x = linear_15_cast_fp16)[name = string("op_266_cast_fp16")];
            tensor<int32, [3]> var_267_perm_0 = const()[name = string("op_267_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_5_axes_0 = const()[name = string("v_5_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_267_cast_fp16 = transpose(perm = var_267_perm_0, x = var_266_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 14, 65, 64]> v_5_cast_fp16 = expand_dims(axes = v_5_axes_0, x = var_267_cast_fp16)[name = string("v_5_cast_fp16")];
            fp16 mul_2_y_0_to_fp16 = const()[name = string("mul_2_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_2_cast_fp16 = mul(x = q_5_cast_fp16, y = mul_2_y_0_to_fp16)[name = string("mul_2_cast_fp16")];
            bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(true)];
            bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_2_cast_fp16, y = k_5_cast_fp16)[name = string("matmul_2_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = attn_mask_to_fp16)[name = string("add_2_cast_fp16")];
            int32 softmax_2_axis_0 = const()[name = string("softmax_2_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16)[name = string("softmax_2_cast_fp16")];
            bool a_5_transpose_x_0 = const()[name = string("a_5_transpose_x_0"), val = bool(false)];
            bool a_5_transpose_y_0 = const()[name = string("a_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_5_cast_fp16 = matmul(transpose_x = a_5_transpose_x_0, transpose_y = a_5_transpose_y_0, x = softmax_2_cast_fp16, y = v_5_cast_fp16)[name = string("a_5_cast_fp16")];
            tensor<int32, [1]> var_270_axes_0 = const()[name = string("op_270_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_270_cast_fp16 = squeeze(axes = var_270_axes_0, x = a_5_cast_fp16)[name = string("op_270_cast_fp16")];
            tensor<int32, [3]> var_271_perm_0 = const()[name = string("op_271_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_272 = const()[name = string("op_272"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_271_cast_fp16 = transpose(perm = var_271_perm_0, x = var_270_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [65, 896]> input_37_cast_fp16 = reshape(shape = var_272, x = var_271_cast_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_2_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32832640))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33635520))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_2_out_bias_to_fp16 = const()[name = string("trunk_layers_2_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33637376)))];
            tensor<fp16, [65, 896]> linear_16_cast_fp16 = linear(bias = trunk_layers_2_out_bias_to_fp16, weight = trunk_layers_2_out_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<fp16, [65, 896]> input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_16_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<int32, [1]> input_41_axes_0 = const()[name = string("input_41_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_2_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_2_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33639232)))];
            tensor<fp16, [896]> trunk_layers_2_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_2_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33641088)))];
            tensor<fp16, [65, 896]> input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = trunk_layers_2_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_2_mlp_ln_weight_to_fp16, x = input_39_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_2_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33642944))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36854272))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_2_fc1_bias_to_fp16 = const()[name = string("trunk_layers_2_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36861504)))];
            tensor<fp16, [65, 3584]> linear_17_cast_fp16 = linear(bias = trunk_layers_2_fc1_bias_to_fp16, weight = trunk_layers_2_fc1_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_17_cast_fp16")];
            string input_43_mode_0 = const()[name = string("input_43_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_43_cast_fp16 = gelu(mode = input_43_mode_0, x = linear_17_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_2_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_2_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36868736))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40080064))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_2_fc2_bias_to_fp16 = const()[name = string("trunk_layers_2_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40081920)))];
            tensor<fp16, [65, 896]> linear_18_cast_fp16 = linear(bias = trunk_layers_2_fc2_bias_to_fp16, weight = trunk_layers_2_fc2_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [65, 896]> input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_18_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<int32, [1]> input_47_axes_0 = const()[name = string("input_47_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_3_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_3_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40083776)))];
            tensor<fp16, [896]> trunk_layers_3_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_3_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40085632)))];
            tensor<fp16, [65, 896]> input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = trunk_layers_3_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_3_attn_ln_weight_to_fp16, x = input_45_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_3_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40087488))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40890368))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_3_q_bias_to_fp16 = const()[name = string("trunk_layers_3_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40892224)))];
            tensor<fp16, [65, 896]> linear_19_cast_fp16 = linear(bias = trunk_layers_3_q_bias_to_fp16, weight = trunk_layers_3_q_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<int32, [3]> var_306 = const()[name = string("op_306"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_307_cast_fp16 = reshape(shape = var_306, x = linear_19_cast_fp16)[name = string("op_307_cast_fp16")];
            tensor<int32, [3]> var_308_perm_0 = const()[name = string("op_308_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_7_axes_0 = const()[name = string("q_7_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_308_cast_fp16 = transpose(perm = var_308_perm_0, x = var_307_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 14, 65, 64]> q_7_cast_fp16 = expand_dims(axes = q_7_axes_0, x = var_308_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_3_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40894080))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41696960))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_3_k_bias_to_fp16 = const()[name = string("trunk_layers_3_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41698816)))];
            tensor<fp16, [65, 896]> linear_20_cast_fp16 = linear(bias = trunk_layers_3_k_bias_to_fp16, weight = trunk_layers_3_k_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [3]> var_313 = const()[name = string("op_313"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_314_cast_fp16 = reshape(shape = var_313, x = linear_20_cast_fp16)[name = string("op_314_cast_fp16")];
            tensor<int32, [3]> var_315_perm_0 = const()[name = string("op_315_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_7_axes_0 = const()[name = string("k_7_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_315_cast_fp16 = transpose(perm = var_315_perm_0, x = var_314_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 14, 65, 64]> k_7_cast_fp16 = expand_dims(axes = k_7_axes_0, x = var_315_cast_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_3_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41700672))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42503552))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_3_v_bias_to_fp16 = const()[name = string("trunk_layers_3_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42505408)))];
            tensor<fp16, [65, 896]> linear_21_cast_fp16 = linear(bias = trunk_layers_3_v_bias_to_fp16, weight = trunk_layers_3_v_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [3]> var_320 = const()[name = string("op_320"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_321_cast_fp16 = reshape(shape = var_320, x = linear_21_cast_fp16)[name = string("op_321_cast_fp16")];
            tensor<int32, [3]> var_322_perm_0 = const()[name = string("op_322_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_7_axes_0 = const()[name = string("v_7_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_322_cast_fp16 = transpose(perm = var_322_perm_0, x = var_321_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 14, 65, 64]> v_7_cast_fp16 = expand_dims(axes = v_7_axes_0, x = var_322_cast_fp16)[name = string("v_7_cast_fp16")];
            fp16 mul_3_y_0_to_fp16 = const()[name = string("mul_3_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_3_cast_fp16 = mul(x = q_7_cast_fp16, y = mul_3_y_0_to_fp16)[name = string("mul_3_cast_fp16")];
            bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(true)];
            bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_3_cast_fp16, y = k_7_cast_fp16)[name = string("matmul_3_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = attn_mask_to_fp16)[name = string("add_3_cast_fp16")];
            int32 softmax_3_axis_0 = const()[name = string("softmax_3_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = add_3_cast_fp16)[name = string("softmax_3_cast_fp16")];
            bool a_7_transpose_x_0 = const()[name = string("a_7_transpose_x_0"), val = bool(false)];
            bool a_7_transpose_y_0 = const()[name = string("a_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_7_cast_fp16 = matmul(transpose_x = a_7_transpose_x_0, transpose_y = a_7_transpose_y_0, x = softmax_3_cast_fp16, y = v_7_cast_fp16)[name = string("a_7_cast_fp16")];
            tensor<int32, [1]> var_325_axes_0 = const()[name = string("op_325_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_325_cast_fp16 = squeeze(axes = var_325_axes_0, x = a_7_cast_fp16)[name = string("op_325_cast_fp16")];
            tensor<int32, [3]> var_326_perm_0 = const()[name = string("op_326_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_327 = const()[name = string("op_327"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_326_cast_fp16 = transpose(perm = var_326_perm_0, x = var_325_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [65, 896]> input_49_cast_fp16 = reshape(shape = var_327, x = var_326_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_3_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42507264))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43310144))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_3_out_bias_to_fp16 = const()[name = string("trunk_layers_3_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43312000)))];
            tensor<fp16, [65, 896]> linear_22_cast_fp16 = linear(bias = trunk_layers_3_out_bias_to_fp16, weight = trunk_layers_3_out_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<fp16, [65, 896]> input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_22_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<int32, [1]> input_53_axes_0 = const()[name = string("input_53_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_3_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_3_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43313856)))];
            tensor<fp16, [896]> trunk_layers_3_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_3_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43315712)))];
            tensor<fp16, [65, 896]> input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = trunk_layers_3_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_3_mlp_ln_weight_to_fp16, x = input_51_cast_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_3_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43317568))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46528896))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_3_fc1_bias_to_fp16 = const()[name = string("trunk_layers_3_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46536128)))];
            tensor<fp16, [65, 3584]> linear_23_cast_fp16 = linear(bias = trunk_layers_3_fc1_bias_to_fp16, weight = trunk_layers_3_fc1_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_23_cast_fp16")];
            string input_55_mode_0 = const()[name = string("input_55_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_55_cast_fp16 = gelu(mode = input_55_mode_0, x = linear_23_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_3_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_3_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46543360))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49754688))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_3_fc2_bias_to_fp16 = const()[name = string("trunk_layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49756544)))];
            tensor<fp16, [65, 896]> linear_24_cast_fp16 = linear(bias = trunk_layers_3_fc2_bias_to_fp16, weight = trunk_layers_3_fc2_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [65, 896]> input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_24_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<int32, [1]> input_59_axes_0 = const()[name = string("input_59_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_4_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_4_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49758400)))];
            tensor<fp16, [896]> trunk_layers_4_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_4_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49760256)))];
            tensor<fp16, [65, 896]> input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = trunk_layers_4_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_4_attn_ln_weight_to_fp16, x = input_57_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_4_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49762112))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50564992))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_4_q_bias_to_fp16 = const()[name = string("trunk_layers_4_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50566848)))];
            tensor<fp16, [65, 896]> linear_25_cast_fp16 = linear(bias = trunk_layers_4_q_bias_to_fp16, weight = trunk_layers_4_q_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<int32, [3]> var_361 = const()[name = string("op_361"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_362_cast_fp16 = reshape(shape = var_361, x = linear_25_cast_fp16)[name = string("op_362_cast_fp16")];
            tensor<int32, [3]> var_363_perm_0 = const()[name = string("op_363_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_9_axes_0 = const()[name = string("q_9_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_363_cast_fp16 = transpose(perm = var_363_perm_0, x = var_362_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 14, 65, 64]> q_9_cast_fp16 = expand_dims(axes = q_9_axes_0, x = var_363_cast_fp16)[name = string("q_9_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_4_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50568704))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51371584))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_4_k_bias_to_fp16 = const()[name = string("trunk_layers_4_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51373440)))];
            tensor<fp16, [65, 896]> linear_26_cast_fp16 = linear(bias = trunk_layers_4_k_bias_to_fp16, weight = trunk_layers_4_k_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<int32, [3]> var_368 = const()[name = string("op_368"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_369_cast_fp16 = reshape(shape = var_368, x = linear_26_cast_fp16)[name = string("op_369_cast_fp16")];
            tensor<int32, [3]> var_370_perm_0 = const()[name = string("op_370_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_9_axes_0 = const()[name = string("k_9_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_370_cast_fp16 = transpose(perm = var_370_perm_0, x = var_369_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 14, 65, 64]> k_9_cast_fp16 = expand_dims(axes = k_9_axes_0, x = var_370_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_4_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51375296))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52178176))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_4_v_bias_to_fp16 = const()[name = string("trunk_layers_4_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52180032)))];
            tensor<fp16, [65, 896]> linear_27_cast_fp16 = linear(bias = trunk_layers_4_v_bias_to_fp16, weight = trunk_layers_4_v_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<int32, [3]> var_375 = const()[name = string("op_375"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_376_cast_fp16 = reshape(shape = var_375, x = linear_27_cast_fp16)[name = string("op_376_cast_fp16")];
            tensor<int32, [3]> var_377_perm_0 = const()[name = string("op_377_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_9_axes_0 = const()[name = string("v_9_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_377_cast_fp16 = transpose(perm = var_377_perm_0, x = var_376_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 14, 65, 64]> v_9_cast_fp16 = expand_dims(axes = v_9_axes_0, x = var_377_cast_fp16)[name = string("v_9_cast_fp16")];
            fp16 mul_4_y_0_to_fp16 = const()[name = string("mul_4_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_4_cast_fp16 = mul(x = q_9_cast_fp16, y = mul_4_y_0_to_fp16)[name = string("mul_4_cast_fp16")];
            bool matmul_4_transpose_y_0 = const()[name = string("matmul_4_transpose_y_0"), val = bool(true)];
            bool matmul_4_transpose_x_0 = const()[name = string("matmul_4_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_4_cast_fp16, y = k_9_cast_fp16)[name = string("matmul_4_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = attn_mask_to_fp16)[name = string("add_4_cast_fp16")];
            int32 softmax_4_axis_0 = const()[name = string("softmax_4_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = add_4_cast_fp16)[name = string("softmax_4_cast_fp16")];
            bool a_9_transpose_x_0 = const()[name = string("a_9_transpose_x_0"), val = bool(false)];
            bool a_9_transpose_y_0 = const()[name = string("a_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_9_cast_fp16 = matmul(transpose_x = a_9_transpose_x_0, transpose_y = a_9_transpose_y_0, x = softmax_4_cast_fp16, y = v_9_cast_fp16)[name = string("a_9_cast_fp16")];
            tensor<int32, [1]> var_380_axes_0 = const()[name = string("op_380_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_380_cast_fp16 = squeeze(axes = var_380_axes_0, x = a_9_cast_fp16)[name = string("op_380_cast_fp16")];
            tensor<int32, [3]> var_381_perm_0 = const()[name = string("op_381_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_382 = const()[name = string("op_382"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_381_cast_fp16 = transpose(perm = var_381_perm_0, x = var_380_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [65, 896]> input_61_cast_fp16 = reshape(shape = var_382, x = var_381_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_4_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52181888))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52984768))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_4_out_bias_to_fp16 = const()[name = string("trunk_layers_4_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52986624)))];
            tensor<fp16, [65, 896]> linear_28_cast_fp16 = linear(bias = trunk_layers_4_out_bias_to_fp16, weight = trunk_layers_4_out_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<fp16, [65, 896]> input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_28_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<int32, [1]> input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_4_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_4_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52988480)))];
            tensor<fp16, [896]> trunk_layers_4_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_4_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52990336)))];
            tensor<fp16, [65, 896]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = trunk_layers_4_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_4_mlp_ln_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_4_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52992192))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56203520))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_4_fc1_bias_to_fp16 = const()[name = string("trunk_layers_4_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56210752)))];
            tensor<fp16, [65, 3584]> linear_29_cast_fp16 = linear(bias = trunk_layers_4_fc1_bias_to_fp16, weight = trunk_layers_4_fc1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_29_cast_fp16")];
            string input_67_mode_0 = const()[name = string("input_67_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_67_cast_fp16 = gelu(mode = input_67_mode_0, x = linear_29_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_4_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_4_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56217984))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59429312))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_4_fc2_bias_to_fp16 = const()[name = string("trunk_layers_4_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59431168)))];
            tensor<fp16, [65, 896]> linear_30_cast_fp16 = linear(bias = trunk_layers_4_fc2_bias_to_fp16, weight = trunk_layers_4_fc2_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<fp16, [65, 896]> input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_30_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<int32, [1]> input_71_axes_0 = const()[name = string("input_71_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_5_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_5_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59433024)))];
            tensor<fp16, [896]> trunk_layers_5_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_5_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59434880)))];
            tensor<fp16, [65, 896]> input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = trunk_layers_5_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_5_attn_ln_weight_to_fp16, x = input_69_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_5_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59436736))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60239616))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_5_q_bias_to_fp16 = const()[name = string("trunk_layers_5_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60241472)))];
            tensor<fp16, [65, 896]> linear_31_cast_fp16 = linear(bias = trunk_layers_5_q_bias_to_fp16, weight = trunk_layers_5_q_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<int32, [3]> var_416 = const()[name = string("op_416"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_417_cast_fp16 = reshape(shape = var_416, x = linear_31_cast_fp16)[name = string("op_417_cast_fp16")];
            tensor<int32, [3]> var_418_perm_0 = const()[name = string("op_418_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_11_axes_0 = const()[name = string("q_11_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_418_cast_fp16 = transpose(perm = var_418_perm_0, x = var_417_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 14, 65, 64]> q_11_cast_fp16 = expand_dims(axes = q_11_axes_0, x = var_418_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_5_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60243328))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61046208))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_5_k_bias_to_fp16 = const()[name = string("trunk_layers_5_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61048064)))];
            tensor<fp16, [65, 896]> linear_32_cast_fp16 = linear(bias = trunk_layers_5_k_bias_to_fp16, weight = trunk_layers_5_k_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<int32, [3]> var_423 = const()[name = string("op_423"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_424_cast_fp16 = reshape(shape = var_423, x = linear_32_cast_fp16)[name = string("op_424_cast_fp16")];
            tensor<int32, [3]> var_425_perm_0 = const()[name = string("op_425_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_11_axes_0 = const()[name = string("k_11_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_425_cast_fp16 = transpose(perm = var_425_perm_0, x = var_424_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 14, 65, 64]> k_11_cast_fp16 = expand_dims(axes = k_11_axes_0, x = var_425_cast_fp16)[name = string("k_11_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_5_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61049920))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61852800))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_5_v_bias_to_fp16 = const()[name = string("trunk_layers_5_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61854656)))];
            tensor<fp16, [65, 896]> linear_33_cast_fp16 = linear(bias = trunk_layers_5_v_bias_to_fp16, weight = trunk_layers_5_v_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<int32, [3]> var_430 = const()[name = string("op_430"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_431_cast_fp16 = reshape(shape = var_430, x = linear_33_cast_fp16)[name = string("op_431_cast_fp16")];
            tensor<int32, [3]> var_432_perm_0 = const()[name = string("op_432_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_11_axes_0 = const()[name = string("v_11_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_432_cast_fp16 = transpose(perm = var_432_perm_0, x = var_431_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 14, 65, 64]> v_11_cast_fp16 = expand_dims(axes = v_11_axes_0, x = var_432_cast_fp16)[name = string("v_11_cast_fp16")];
            fp16 mul_5_y_0_to_fp16 = const()[name = string("mul_5_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_5_cast_fp16 = mul(x = q_11_cast_fp16, y = mul_5_y_0_to_fp16)[name = string("mul_5_cast_fp16")];
            bool matmul_5_transpose_y_0 = const()[name = string("matmul_5_transpose_y_0"), val = bool(true)];
            bool matmul_5_transpose_x_0 = const()[name = string("matmul_5_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_5_cast_fp16, y = k_11_cast_fp16)[name = string("matmul_5_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = attn_mask_to_fp16)[name = string("add_5_cast_fp16")];
            int32 softmax_5_axis_0 = const()[name = string("softmax_5_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = add_5_cast_fp16)[name = string("softmax_5_cast_fp16")];
            bool a_11_transpose_x_0 = const()[name = string("a_11_transpose_x_0"), val = bool(false)];
            bool a_11_transpose_y_0 = const()[name = string("a_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_11_cast_fp16 = matmul(transpose_x = a_11_transpose_x_0, transpose_y = a_11_transpose_y_0, x = softmax_5_cast_fp16, y = v_11_cast_fp16)[name = string("a_11_cast_fp16")];
            tensor<int32, [1]> var_435_axes_0 = const()[name = string("op_435_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_435_cast_fp16 = squeeze(axes = var_435_axes_0, x = a_11_cast_fp16)[name = string("op_435_cast_fp16")];
            tensor<int32, [3]> var_436_perm_0 = const()[name = string("op_436_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_437 = const()[name = string("op_437"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_436_cast_fp16 = transpose(perm = var_436_perm_0, x = var_435_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [65, 896]> input_73_cast_fp16 = reshape(shape = var_437, x = var_436_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_5_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61856512))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62659392))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_5_out_bias_to_fp16 = const()[name = string("trunk_layers_5_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62661248)))];
            tensor<fp16, [65, 896]> linear_34_cast_fp16 = linear(bias = trunk_layers_5_out_bias_to_fp16, weight = trunk_layers_5_out_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [65, 896]> input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_34_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<int32, [1]> input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_5_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_5_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62663104)))];
            tensor<fp16, [896]> trunk_layers_5_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_5_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62664960)))];
            tensor<fp16, [65, 896]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = trunk_layers_5_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_5_mlp_ln_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_5_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62666816))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65878144))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_5_fc1_bias_to_fp16 = const()[name = string("trunk_layers_5_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65885376)))];
            tensor<fp16, [65, 3584]> linear_35_cast_fp16 = linear(bias = trunk_layers_5_fc1_bias_to_fp16, weight = trunk_layers_5_fc1_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_35_cast_fp16")];
            string input_79_mode_0 = const()[name = string("input_79_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = linear_35_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_5_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_5_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65892608))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69103936))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_5_fc2_bias_to_fp16 = const()[name = string("trunk_layers_5_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69105792)))];
            tensor<fp16, [65, 896]> linear_36_cast_fp16 = linear(bias = trunk_layers_5_fc2_bias_to_fp16, weight = trunk_layers_5_fc2_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<fp16, [65, 896]> input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_36_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<int32, [1]> input_83_axes_0 = const()[name = string("input_83_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_6_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_6_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69107648)))];
            tensor<fp16, [896]> trunk_layers_6_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_6_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69109504)))];
            tensor<fp16, [65, 896]> input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = trunk_layers_6_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_6_attn_ln_weight_to_fp16, x = input_81_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_6_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69111360))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69914240))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_6_q_bias_to_fp16 = const()[name = string("trunk_layers_6_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69916096)))];
            tensor<fp16, [65, 896]> linear_37_cast_fp16 = linear(bias = trunk_layers_6_q_bias_to_fp16, weight = trunk_layers_6_q_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [3]> var_471 = const()[name = string("op_471"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_472_cast_fp16 = reshape(shape = var_471, x = linear_37_cast_fp16)[name = string("op_472_cast_fp16")];
            tensor<int32, [3]> var_473_perm_0 = const()[name = string("op_473_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_13_axes_0 = const()[name = string("q_13_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_473_cast_fp16 = transpose(perm = var_473_perm_0, x = var_472_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 14, 65, 64]> q_13_cast_fp16 = expand_dims(axes = q_13_axes_0, x = var_473_cast_fp16)[name = string("q_13_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_6_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69917952))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70720832))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_6_k_bias_to_fp16 = const()[name = string("trunk_layers_6_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70722688)))];
            tensor<fp16, [65, 896]> linear_38_cast_fp16 = linear(bias = trunk_layers_6_k_bias_to_fp16, weight = trunk_layers_6_k_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [3]> var_478 = const()[name = string("op_478"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_479_cast_fp16 = reshape(shape = var_478, x = linear_38_cast_fp16)[name = string("op_479_cast_fp16")];
            tensor<int32, [3]> var_480_perm_0 = const()[name = string("op_480_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_13_axes_0 = const()[name = string("k_13_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_480_cast_fp16 = transpose(perm = var_480_perm_0, x = var_479_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 14, 65, 64]> k_13_cast_fp16 = expand_dims(axes = k_13_axes_0, x = var_480_cast_fp16)[name = string("k_13_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_6_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70724544))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71527424))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_6_v_bias_to_fp16 = const()[name = string("trunk_layers_6_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71529280)))];
            tensor<fp16, [65, 896]> linear_39_cast_fp16 = linear(bias = trunk_layers_6_v_bias_to_fp16, weight = trunk_layers_6_v_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<int32, [3]> var_485 = const()[name = string("op_485"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_486_cast_fp16 = reshape(shape = var_485, x = linear_39_cast_fp16)[name = string("op_486_cast_fp16")];
            tensor<int32, [3]> var_487_perm_0 = const()[name = string("op_487_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_13_axes_0 = const()[name = string("v_13_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_487_cast_fp16 = transpose(perm = var_487_perm_0, x = var_486_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 14, 65, 64]> v_13_cast_fp16 = expand_dims(axes = v_13_axes_0, x = var_487_cast_fp16)[name = string("v_13_cast_fp16")];
            fp16 mul_6_y_0_to_fp16 = const()[name = string("mul_6_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_6_cast_fp16 = mul(x = q_13_cast_fp16, y = mul_6_y_0_to_fp16)[name = string("mul_6_cast_fp16")];
            bool matmul_6_transpose_y_0 = const()[name = string("matmul_6_transpose_y_0"), val = bool(true)];
            bool matmul_6_transpose_x_0 = const()[name = string("matmul_6_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_6_cast_fp16, y = k_13_cast_fp16)[name = string("matmul_6_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_6_cast_fp16 = add(x = matmul_6_cast_fp16, y = attn_mask_to_fp16)[name = string("add_6_cast_fp16")];
            int32 softmax_6_axis_0 = const()[name = string("softmax_6_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = add_6_cast_fp16)[name = string("softmax_6_cast_fp16")];
            bool a_13_transpose_x_0 = const()[name = string("a_13_transpose_x_0"), val = bool(false)];
            bool a_13_transpose_y_0 = const()[name = string("a_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_13_cast_fp16 = matmul(transpose_x = a_13_transpose_x_0, transpose_y = a_13_transpose_y_0, x = softmax_6_cast_fp16, y = v_13_cast_fp16)[name = string("a_13_cast_fp16")];
            tensor<int32, [1]> var_490_axes_0 = const()[name = string("op_490_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_490_cast_fp16 = squeeze(axes = var_490_axes_0, x = a_13_cast_fp16)[name = string("op_490_cast_fp16")];
            tensor<int32, [3]> var_491_perm_0 = const()[name = string("op_491_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_492 = const()[name = string("op_492"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_491_cast_fp16 = transpose(perm = var_491_perm_0, x = var_490_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [65, 896]> input_85_cast_fp16 = reshape(shape = var_492, x = var_491_cast_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_6_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71531136))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72334016))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_6_out_bias_to_fp16 = const()[name = string("trunk_layers_6_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72335872)))];
            tensor<fp16, [65, 896]> linear_40_cast_fp16 = linear(bias = trunk_layers_6_out_bias_to_fp16, weight = trunk_layers_6_out_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [65, 896]> input_87_cast_fp16 = add(x = input_81_cast_fp16, y = linear_40_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<int32, [1]> input_89_axes_0 = const()[name = string("input_89_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_6_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_6_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72337728)))];
            tensor<fp16, [896]> trunk_layers_6_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_6_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72339584)))];
            tensor<fp16, [65, 896]> input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = trunk_layers_6_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_6_mlp_ln_weight_to_fp16, x = input_87_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_6_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72341440))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75552768))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_6_fc1_bias_to_fp16 = const()[name = string("trunk_layers_6_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75560000)))];
            tensor<fp16, [65, 3584]> linear_41_cast_fp16 = linear(bias = trunk_layers_6_fc1_bias_to_fp16, weight = trunk_layers_6_fc1_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_41_cast_fp16")];
            string input_91_mode_0 = const()[name = string("input_91_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = linear_41_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_6_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_6_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75567232))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78778560))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_6_fc2_bias_to_fp16 = const()[name = string("trunk_layers_6_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78780416)))];
            tensor<fp16, [65, 896]> linear_42_cast_fp16 = linear(bias = trunk_layers_6_fc2_bias_to_fp16, weight = trunk_layers_6_fc2_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [65, 896]> input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_42_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<int32, [1]> input_95_axes_0 = const()[name = string("input_95_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_7_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_7_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78782272)))];
            tensor<fp16, [896]> trunk_layers_7_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_7_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78784128)))];
            tensor<fp16, [65, 896]> input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = trunk_layers_7_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_7_attn_ln_weight_to_fp16, x = input_93_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_7_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78785984))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79588864))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_7_q_bias_to_fp16 = const()[name = string("trunk_layers_7_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79590720)))];
            tensor<fp16, [65, 896]> linear_43_cast_fp16 = linear(bias = trunk_layers_7_q_bias_to_fp16, weight = trunk_layers_7_q_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [3]> var_526 = const()[name = string("op_526"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_527_cast_fp16 = reshape(shape = var_526, x = linear_43_cast_fp16)[name = string("op_527_cast_fp16")];
            tensor<int32, [3]> var_528_perm_0 = const()[name = string("op_528_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_15_axes_0 = const()[name = string("q_15_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_528_cast_fp16 = transpose(perm = var_528_perm_0, x = var_527_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 14, 65, 64]> q_15_cast_fp16 = expand_dims(axes = q_15_axes_0, x = var_528_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_7_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79592576))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80395456))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_7_k_bias_to_fp16 = const()[name = string("trunk_layers_7_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80397312)))];
            tensor<fp16, [65, 896]> linear_44_cast_fp16 = linear(bias = trunk_layers_7_k_bias_to_fp16, weight = trunk_layers_7_k_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [3]> var_533 = const()[name = string("op_533"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_534_cast_fp16 = reshape(shape = var_533, x = linear_44_cast_fp16)[name = string("op_534_cast_fp16")];
            tensor<int32, [3]> var_535_perm_0 = const()[name = string("op_535_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_15_axes_0 = const()[name = string("k_15_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_535_cast_fp16 = transpose(perm = var_535_perm_0, x = var_534_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 14, 65, 64]> k_15_cast_fp16 = expand_dims(axes = k_15_axes_0, x = var_535_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_7_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80399168))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81202048))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_7_v_bias_to_fp16 = const()[name = string("trunk_layers_7_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81203904)))];
            tensor<fp16, [65, 896]> linear_45_cast_fp16 = linear(bias = trunk_layers_7_v_bias_to_fp16, weight = trunk_layers_7_v_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<int32, [3]> var_540 = const()[name = string("op_540"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_541_cast_fp16 = reshape(shape = var_540, x = linear_45_cast_fp16)[name = string("op_541_cast_fp16")];
            tensor<int32, [3]> var_542_perm_0 = const()[name = string("op_542_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_15_axes_0 = const()[name = string("v_15_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_542_cast_fp16 = transpose(perm = var_542_perm_0, x = var_541_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 14, 65, 64]> v_15_cast_fp16 = expand_dims(axes = v_15_axes_0, x = var_542_cast_fp16)[name = string("v_15_cast_fp16")];
            fp16 mul_7_y_0_to_fp16 = const()[name = string("mul_7_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_7_cast_fp16 = mul(x = q_15_cast_fp16, y = mul_7_y_0_to_fp16)[name = string("mul_7_cast_fp16")];
            bool matmul_7_transpose_y_0 = const()[name = string("matmul_7_transpose_y_0"), val = bool(true)];
            bool matmul_7_transpose_x_0 = const()[name = string("matmul_7_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_7_cast_fp16, y = k_15_cast_fp16)[name = string("matmul_7_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_7_cast_fp16 = add(x = matmul_7_cast_fp16, y = attn_mask_to_fp16)[name = string("add_7_cast_fp16")];
            int32 softmax_7_axis_0 = const()[name = string("softmax_7_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = add_7_cast_fp16)[name = string("softmax_7_cast_fp16")];
            bool a_15_transpose_x_0 = const()[name = string("a_15_transpose_x_0"), val = bool(false)];
            bool a_15_transpose_y_0 = const()[name = string("a_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_15_cast_fp16 = matmul(transpose_x = a_15_transpose_x_0, transpose_y = a_15_transpose_y_0, x = softmax_7_cast_fp16, y = v_15_cast_fp16)[name = string("a_15_cast_fp16")];
            tensor<int32, [1]> var_545_axes_0 = const()[name = string("op_545_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_545_cast_fp16 = squeeze(axes = var_545_axes_0, x = a_15_cast_fp16)[name = string("op_545_cast_fp16")];
            tensor<int32, [3]> var_546_perm_0 = const()[name = string("op_546_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_547 = const()[name = string("op_547"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_546_cast_fp16 = transpose(perm = var_546_perm_0, x = var_545_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [65, 896]> input_97_cast_fp16 = reshape(shape = var_547, x = var_546_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_7_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81205760))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82008640))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_7_out_bias_to_fp16 = const()[name = string("trunk_layers_7_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82010496)))];
            tensor<fp16, [65, 896]> linear_46_cast_fp16 = linear(bias = trunk_layers_7_out_bias_to_fp16, weight = trunk_layers_7_out_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [65, 896]> input_99_cast_fp16 = add(x = input_93_cast_fp16, y = linear_46_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<int32, [1]> input_101_axes_0 = const()[name = string("input_101_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_7_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_7_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82012352)))];
            tensor<fp16, [896]> trunk_layers_7_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_7_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82014208)))];
            tensor<fp16, [65, 896]> input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = trunk_layers_7_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_7_mlp_ln_weight_to_fp16, x = input_99_cast_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_7_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82016064))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85227392))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_7_fc1_bias_to_fp16 = const()[name = string("trunk_layers_7_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85234624)))];
            tensor<fp16, [65, 3584]> linear_47_cast_fp16 = linear(bias = trunk_layers_7_fc1_bias_to_fp16, weight = trunk_layers_7_fc1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_47_cast_fp16")];
            string input_103_mode_0 = const()[name = string("input_103_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_103_cast_fp16 = gelu(mode = input_103_mode_0, x = linear_47_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_7_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_7_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85241856))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88453184))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_7_fc2_bias_to_fp16 = const()[name = string("trunk_layers_7_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88455040)))];
            tensor<fp16, [65, 896]> linear_48_cast_fp16 = linear(bias = trunk_layers_7_fc2_bias_to_fp16, weight = trunk_layers_7_fc2_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [65, 896]> input_105_cast_fp16 = add(x = input_99_cast_fp16, y = linear_48_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<int32, [1]> input_107_axes_0 = const()[name = string("input_107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_8_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_8_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88456896)))];
            tensor<fp16, [896]> trunk_layers_8_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_8_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88458752)))];
            tensor<fp16, [65, 896]> input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = trunk_layers_8_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_8_attn_ln_weight_to_fp16, x = input_105_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_8_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88460608))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89263488))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_8_q_bias_to_fp16 = const()[name = string("trunk_layers_8_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89265344)))];
            tensor<fp16, [65, 896]> linear_49_cast_fp16 = linear(bias = trunk_layers_8_q_bias_to_fp16, weight = trunk_layers_8_q_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [3]> var_581 = const()[name = string("op_581"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_582_cast_fp16 = reshape(shape = var_581, x = linear_49_cast_fp16)[name = string("op_582_cast_fp16")];
            tensor<int32, [3]> var_583_perm_0 = const()[name = string("op_583_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_17_axes_0 = const()[name = string("q_17_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_583_cast_fp16 = transpose(perm = var_583_perm_0, x = var_582_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 14, 65, 64]> q_17_cast_fp16 = expand_dims(axes = q_17_axes_0, x = var_583_cast_fp16)[name = string("q_17_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_8_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89267200))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90070080))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_8_k_bias_to_fp16 = const()[name = string("trunk_layers_8_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90071936)))];
            tensor<fp16, [65, 896]> linear_50_cast_fp16 = linear(bias = trunk_layers_8_k_bias_to_fp16, weight = trunk_layers_8_k_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [3]> var_588 = const()[name = string("op_588"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_589_cast_fp16 = reshape(shape = var_588, x = linear_50_cast_fp16)[name = string("op_589_cast_fp16")];
            tensor<int32, [3]> var_590_perm_0 = const()[name = string("op_590_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_17_axes_0 = const()[name = string("k_17_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_590_cast_fp16 = transpose(perm = var_590_perm_0, x = var_589_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 14, 65, 64]> k_17_cast_fp16 = expand_dims(axes = k_17_axes_0, x = var_590_cast_fp16)[name = string("k_17_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_8_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90073792))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90876672))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_8_v_bias_to_fp16 = const()[name = string("trunk_layers_8_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90878528)))];
            tensor<fp16, [65, 896]> linear_51_cast_fp16 = linear(bias = trunk_layers_8_v_bias_to_fp16, weight = trunk_layers_8_v_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [3]> var_595 = const()[name = string("op_595"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_596_cast_fp16 = reshape(shape = var_595, x = linear_51_cast_fp16)[name = string("op_596_cast_fp16")];
            tensor<int32, [3]> var_597_perm_0 = const()[name = string("op_597_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_17_axes_0 = const()[name = string("v_17_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_597_cast_fp16 = transpose(perm = var_597_perm_0, x = var_596_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 14, 65, 64]> v_17_cast_fp16 = expand_dims(axes = v_17_axes_0, x = var_597_cast_fp16)[name = string("v_17_cast_fp16")];
            fp16 mul_8_y_0_to_fp16 = const()[name = string("mul_8_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_8_cast_fp16 = mul(x = q_17_cast_fp16, y = mul_8_y_0_to_fp16)[name = string("mul_8_cast_fp16")];
            bool matmul_8_transpose_y_0 = const()[name = string("matmul_8_transpose_y_0"), val = bool(true)];
            bool matmul_8_transpose_x_0 = const()[name = string("matmul_8_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_8_cast_fp16, y = k_17_cast_fp16)[name = string("matmul_8_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_8_cast_fp16 = add(x = matmul_8_cast_fp16, y = attn_mask_to_fp16)[name = string("add_8_cast_fp16")];
            int32 softmax_8_axis_0 = const()[name = string("softmax_8_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = add_8_cast_fp16)[name = string("softmax_8_cast_fp16")];
            bool a_17_transpose_x_0 = const()[name = string("a_17_transpose_x_0"), val = bool(false)];
            bool a_17_transpose_y_0 = const()[name = string("a_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_17_cast_fp16 = matmul(transpose_x = a_17_transpose_x_0, transpose_y = a_17_transpose_y_0, x = softmax_8_cast_fp16, y = v_17_cast_fp16)[name = string("a_17_cast_fp16")];
            tensor<int32, [1]> var_600_axes_0 = const()[name = string("op_600_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_600_cast_fp16 = squeeze(axes = var_600_axes_0, x = a_17_cast_fp16)[name = string("op_600_cast_fp16")];
            tensor<int32, [3]> var_601_perm_0 = const()[name = string("op_601_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_602 = const()[name = string("op_602"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_601_cast_fp16 = transpose(perm = var_601_perm_0, x = var_600_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [65, 896]> input_109_cast_fp16 = reshape(shape = var_602, x = var_601_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_8_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90880384))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91683264))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_8_out_bias_to_fp16 = const()[name = string("trunk_layers_8_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91685120)))];
            tensor<fp16, [65, 896]> linear_52_cast_fp16 = linear(bias = trunk_layers_8_out_bias_to_fp16, weight = trunk_layers_8_out_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [65, 896]> input_111_cast_fp16 = add(x = input_105_cast_fp16, y = linear_52_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<int32, [1]> input_113_axes_0 = const()[name = string("input_113_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_8_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_8_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91686976)))];
            tensor<fp16, [896]> trunk_layers_8_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_8_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91688832)))];
            tensor<fp16, [65, 896]> input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = trunk_layers_8_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_8_mlp_ln_weight_to_fp16, x = input_111_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_8_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91690688))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94902016))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_8_fc1_bias_to_fp16 = const()[name = string("trunk_layers_8_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94909248)))];
            tensor<fp16, [65, 3584]> linear_53_cast_fp16 = linear(bias = trunk_layers_8_fc1_bias_to_fp16, weight = trunk_layers_8_fc1_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_53_cast_fp16")];
            string input_115_mode_0 = const()[name = string("input_115_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_115_cast_fp16 = gelu(mode = input_115_mode_0, x = linear_53_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_8_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_8_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94916480))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98127808))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_8_fc2_bias_to_fp16 = const()[name = string("trunk_layers_8_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98129664)))];
            tensor<fp16, [65, 896]> linear_54_cast_fp16 = linear(bias = trunk_layers_8_fc2_bias_to_fp16, weight = trunk_layers_8_fc2_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [65, 896]> input_117_cast_fp16 = add(x = input_111_cast_fp16, y = linear_54_cast_fp16)[name = string("input_117_cast_fp16")];
            tensor<int32, [1]> input_119_axes_0 = const()[name = string("input_119_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_9_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_9_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98131520)))];
            tensor<fp16, [896]> trunk_layers_9_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_9_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98133376)))];
            tensor<fp16, [65, 896]> input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = trunk_layers_9_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_9_attn_ln_weight_to_fp16, x = input_117_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_9_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98135232))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98938112))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_9_q_bias_to_fp16 = const()[name = string("trunk_layers_9_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98939968)))];
            tensor<fp16, [65, 896]> linear_55_cast_fp16 = linear(bias = trunk_layers_9_q_bias_to_fp16, weight = trunk_layers_9_q_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<int32, [3]> var_636 = const()[name = string("op_636"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_637_cast_fp16 = reshape(shape = var_636, x = linear_55_cast_fp16)[name = string("op_637_cast_fp16")];
            tensor<int32, [3]> var_638_perm_0 = const()[name = string("op_638_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_19_axes_0 = const()[name = string("q_19_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_638_cast_fp16 = transpose(perm = var_638_perm_0, x = var_637_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 14, 65, 64]> q_19_cast_fp16 = expand_dims(axes = q_19_axes_0, x = var_638_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_9_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98941824))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99744704))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_9_k_bias_to_fp16 = const()[name = string("trunk_layers_9_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99746560)))];
            tensor<fp16, [65, 896]> linear_56_cast_fp16 = linear(bias = trunk_layers_9_k_bias_to_fp16, weight = trunk_layers_9_k_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [3]> var_643 = const()[name = string("op_643"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_644_cast_fp16 = reshape(shape = var_643, x = linear_56_cast_fp16)[name = string("op_644_cast_fp16")];
            tensor<int32, [3]> var_645_perm_0 = const()[name = string("op_645_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_19_axes_0 = const()[name = string("k_19_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_645_cast_fp16 = transpose(perm = var_645_perm_0, x = var_644_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 14, 65, 64]> k_19_cast_fp16 = expand_dims(axes = k_19_axes_0, x = var_645_cast_fp16)[name = string("k_19_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_9_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99748416))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100551296))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_9_v_bias_to_fp16 = const()[name = string("trunk_layers_9_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100553152)))];
            tensor<fp16, [65, 896]> linear_57_cast_fp16 = linear(bias = trunk_layers_9_v_bias_to_fp16, weight = trunk_layers_9_v_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [3]> var_650 = const()[name = string("op_650"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_651_cast_fp16 = reshape(shape = var_650, x = linear_57_cast_fp16)[name = string("op_651_cast_fp16")];
            tensor<int32, [3]> var_652_perm_0 = const()[name = string("op_652_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_19_axes_0 = const()[name = string("v_19_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_652_cast_fp16 = transpose(perm = var_652_perm_0, x = var_651_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 14, 65, 64]> v_19_cast_fp16 = expand_dims(axes = v_19_axes_0, x = var_652_cast_fp16)[name = string("v_19_cast_fp16")];
            fp16 mul_9_y_0_to_fp16 = const()[name = string("mul_9_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_9_cast_fp16 = mul(x = q_19_cast_fp16, y = mul_9_y_0_to_fp16)[name = string("mul_9_cast_fp16")];
            bool matmul_9_transpose_y_0 = const()[name = string("matmul_9_transpose_y_0"), val = bool(true)];
            bool matmul_9_transpose_x_0 = const()[name = string("matmul_9_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_9_cast_fp16, y = k_19_cast_fp16)[name = string("matmul_9_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_9_cast_fp16 = add(x = matmul_9_cast_fp16, y = attn_mask_to_fp16)[name = string("add_9_cast_fp16")];
            int32 softmax_9_axis_0 = const()[name = string("softmax_9_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = add_9_cast_fp16)[name = string("softmax_9_cast_fp16")];
            bool a_19_transpose_x_0 = const()[name = string("a_19_transpose_x_0"), val = bool(false)];
            bool a_19_transpose_y_0 = const()[name = string("a_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_19_cast_fp16 = matmul(transpose_x = a_19_transpose_x_0, transpose_y = a_19_transpose_y_0, x = softmax_9_cast_fp16, y = v_19_cast_fp16)[name = string("a_19_cast_fp16")];
            tensor<int32, [1]> var_655_axes_0 = const()[name = string("op_655_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_655_cast_fp16 = squeeze(axes = var_655_axes_0, x = a_19_cast_fp16)[name = string("op_655_cast_fp16")];
            tensor<int32, [3]> var_656_perm_0 = const()[name = string("op_656_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_657 = const()[name = string("op_657"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_656_cast_fp16 = transpose(perm = var_656_perm_0, x = var_655_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [65, 896]> input_121_cast_fp16 = reshape(shape = var_657, x = var_656_cast_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_9_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100555008))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101357888))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_9_out_bias_to_fp16 = const()[name = string("trunk_layers_9_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101359744)))];
            tensor<fp16, [65, 896]> linear_58_cast_fp16 = linear(bias = trunk_layers_9_out_bias_to_fp16, weight = trunk_layers_9_out_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<fp16, [65, 896]> input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_58_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<int32, [1]> input_125_axes_0 = const()[name = string("input_125_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_9_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_9_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101361600)))];
            tensor<fp16, [896]> trunk_layers_9_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_9_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101363456)))];
            tensor<fp16, [65, 896]> input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = trunk_layers_9_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_9_mlp_ln_weight_to_fp16, x = input_123_cast_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_9_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101365312))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104576640))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_9_fc1_bias_to_fp16 = const()[name = string("trunk_layers_9_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104583872)))];
            tensor<fp16, [65, 3584]> linear_59_cast_fp16 = linear(bias = trunk_layers_9_fc1_bias_to_fp16, weight = trunk_layers_9_fc1_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_59_cast_fp16")];
            string input_127_mode_0 = const()[name = string("input_127_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_127_cast_fp16 = gelu(mode = input_127_mode_0, x = linear_59_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_9_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_9_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104591104))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107802432))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_9_fc2_bias_to_fp16 = const()[name = string("trunk_layers_9_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107804288)))];
            tensor<fp16, [65, 896]> linear_60_cast_fp16 = linear(bias = trunk_layers_9_fc2_bias_to_fp16, weight = trunk_layers_9_fc2_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [65, 896]> input_129_cast_fp16 = add(x = input_123_cast_fp16, y = linear_60_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<int32, [1]> input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_10_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_10_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107806144)))];
            tensor<fp16, [896]> trunk_layers_10_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_10_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107808000)))];
            tensor<fp16, [65, 896]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = trunk_layers_10_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_10_attn_ln_weight_to_fp16, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_10_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107809856))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108612736))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_10_q_bias_to_fp16 = const()[name = string("trunk_layers_10_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108614592)))];
            tensor<fp16, [65, 896]> linear_61_cast_fp16 = linear(bias = trunk_layers_10_q_bias_to_fp16, weight = trunk_layers_10_q_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<int32, [3]> var_691 = const()[name = string("op_691"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_692_cast_fp16 = reshape(shape = var_691, x = linear_61_cast_fp16)[name = string("op_692_cast_fp16")];
            tensor<int32, [3]> var_693_perm_0 = const()[name = string("op_693_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_21_axes_0 = const()[name = string("q_21_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_693_cast_fp16 = transpose(perm = var_693_perm_0, x = var_692_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 14, 65, 64]> q_21_cast_fp16 = expand_dims(axes = q_21_axes_0, x = var_693_cast_fp16)[name = string("q_21_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_10_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108616448))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109419328))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_10_k_bias_to_fp16 = const()[name = string("trunk_layers_10_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109421184)))];
            tensor<fp16, [65, 896]> linear_62_cast_fp16 = linear(bias = trunk_layers_10_k_bias_to_fp16, weight = trunk_layers_10_k_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<int32, [3]> var_698 = const()[name = string("op_698"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_699_cast_fp16 = reshape(shape = var_698, x = linear_62_cast_fp16)[name = string("op_699_cast_fp16")];
            tensor<int32, [3]> var_700_perm_0 = const()[name = string("op_700_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_21_axes_0 = const()[name = string("k_21_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_700_cast_fp16 = transpose(perm = var_700_perm_0, x = var_699_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 14, 65, 64]> k_21_cast_fp16 = expand_dims(axes = k_21_axes_0, x = var_700_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_10_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109423040))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110225920))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_10_v_bias_to_fp16 = const()[name = string("trunk_layers_10_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110227776)))];
            tensor<fp16, [65, 896]> linear_63_cast_fp16 = linear(bias = trunk_layers_10_v_bias_to_fp16, weight = trunk_layers_10_v_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [3]> var_705 = const()[name = string("op_705"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_706_cast_fp16 = reshape(shape = var_705, x = linear_63_cast_fp16)[name = string("op_706_cast_fp16")];
            tensor<int32, [3]> var_707_perm_0 = const()[name = string("op_707_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_21_axes_0 = const()[name = string("v_21_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_707_cast_fp16 = transpose(perm = var_707_perm_0, x = var_706_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 14, 65, 64]> v_21_cast_fp16 = expand_dims(axes = v_21_axes_0, x = var_707_cast_fp16)[name = string("v_21_cast_fp16")];
            fp16 mul_10_y_0_to_fp16 = const()[name = string("mul_10_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_10_cast_fp16 = mul(x = q_21_cast_fp16, y = mul_10_y_0_to_fp16)[name = string("mul_10_cast_fp16")];
            bool matmul_10_transpose_y_0 = const()[name = string("matmul_10_transpose_y_0"), val = bool(true)];
            bool matmul_10_transpose_x_0 = const()[name = string("matmul_10_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_10_cast_fp16, y = k_21_cast_fp16)[name = string("matmul_10_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_10_cast_fp16 = add(x = matmul_10_cast_fp16, y = attn_mask_to_fp16)[name = string("add_10_cast_fp16")];
            int32 softmax_10_axis_0 = const()[name = string("softmax_10_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = add_10_cast_fp16)[name = string("softmax_10_cast_fp16")];
            bool a_21_transpose_x_0 = const()[name = string("a_21_transpose_x_0"), val = bool(false)];
            bool a_21_transpose_y_0 = const()[name = string("a_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_21_cast_fp16 = matmul(transpose_x = a_21_transpose_x_0, transpose_y = a_21_transpose_y_0, x = softmax_10_cast_fp16, y = v_21_cast_fp16)[name = string("a_21_cast_fp16")];
            tensor<int32, [1]> var_710_axes_0 = const()[name = string("op_710_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_710_cast_fp16 = squeeze(axes = var_710_axes_0, x = a_21_cast_fp16)[name = string("op_710_cast_fp16")];
            tensor<int32, [3]> var_711_perm_0 = const()[name = string("op_711_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_712 = const()[name = string("op_712"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_711_cast_fp16 = transpose(perm = var_711_perm_0, x = var_710_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [65, 896]> input_133_cast_fp16 = reshape(shape = var_712, x = var_711_cast_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_10_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110229632))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111032512))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_10_out_bias_to_fp16 = const()[name = string("trunk_layers_10_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111034368)))];
            tensor<fp16, [65, 896]> linear_64_cast_fp16 = linear(bias = trunk_layers_10_out_bias_to_fp16, weight = trunk_layers_10_out_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<fp16, [65, 896]> input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_64_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<int32, [1]> input_137_axes_0 = const()[name = string("input_137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_10_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_10_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111036224)))];
            tensor<fp16, [896]> trunk_layers_10_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_10_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111038080)))];
            tensor<fp16, [65, 896]> input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = trunk_layers_10_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_10_mlp_ln_weight_to_fp16, x = input_135_cast_fp16)[name = string("input_137_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_10_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111039936))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114251264))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_10_fc1_bias_to_fp16 = const()[name = string("trunk_layers_10_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114258496)))];
            tensor<fp16, [65, 3584]> linear_65_cast_fp16 = linear(bias = trunk_layers_10_fc1_bias_to_fp16, weight = trunk_layers_10_fc1_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_65_cast_fp16")];
            string input_139_mode_0 = const()[name = string("input_139_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_139_cast_fp16 = gelu(mode = input_139_mode_0, x = linear_65_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_10_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_10_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114265728))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117477056))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_10_fc2_bias_to_fp16 = const()[name = string("trunk_layers_10_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117478912)))];
            tensor<fp16, [65, 896]> linear_66_cast_fp16 = linear(bias = trunk_layers_10_fc2_bias_to_fp16, weight = trunk_layers_10_fc2_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [65, 896]> input_141_cast_fp16 = add(x = input_135_cast_fp16, y = linear_66_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<int32, [1]> input_143_axes_0 = const()[name = string("input_143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_11_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_11_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117480768)))];
            tensor<fp16, [896]> trunk_layers_11_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_11_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117482624)))];
            tensor<fp16, [65, 896]> input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = trunk_layers_11_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_11_attn_ln_weight_to_fp16, x = input_141_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_11_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117484480))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118287360))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_11_q_bias_to_fp16 = const()[name = string("trunk_layers_11_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118289216)))];
            tensor<fp16, [65, 896]> linear_67_cast_fp16 = linear(bias = trunk_layers_11_q_bias_to_fp16, weight = trunk_layers_11_q_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<int32, [3]> var_746 = const()[name = string("op_746"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_747_cast_fp16 = reshape(shape = var_746, x = linear_67_cast_fp16)[name = string("op_747_cast_fp16")];
            tensor<int32, [3]> var_748_perm_0 = const()[name = string("op_748_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_23_axes_0 = const()[name = string("q_23_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_748_cast_fp16 = transpose(perm = var_748_perm_0, x = var_747_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 14, 65, 64]> q_23_cast_fp16 = expand_dims(axes = q_23_axes_0, x = var_748_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_11_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118291072))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119093952))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_11_k_bias_to_fp16 = const()[name = string("trunk_layers_11_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119095808)))];
            tensor<fp16, [65, 896]> linear_68_cast_fp16 = linear(bias = trunk_layers_11_k_bias_to_fp16, weight = trunk_layers_11_k_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<int32, [3]> var_753 = const()[name = string("op_753"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_754_cast_fp16 = reshape(shape = var_753, x = linear_68_cast_fp16)[name = string("op_754_cast_fp16")];
            tensor<int32, [3]> var_755_perm_0 = const()[name = string("op_755_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_23_axes_0 = const()[name = string("k_23_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_755_cast_fp16 = transpose(perm = var_755_perm_0, x = var_754_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 14, 65, 64]> k_23_cast_fp16 = expand_dims(axes = k_23_axes_0, x = var_755_cast_fp16)[name = string("k_23_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_11_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119097664))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119900544))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_11_v_bias_to_fp16 = const()[name = string("trunk_layers_11_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119902400)))];
            tensor<fp16, [65, 896]> linear_69_cast_fp16 = linear(bias = trunk_layers_11_v_bias_to_fp16, weight = trunk_layers_11_v_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<int32, [3]> var_760 = const()[name = string("op_760"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_761_cast_fp16 = reshape(shape = var_760, x = linear_69_cast_fp16)[name = string("op_761_cast_fp16")];
            tensor<int32, [3]> var_762_perm_0 = const()[name = string("op_762_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_23_axes_0 = const()[name = string("v_23_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_762_cast_fp16 = transpose(perm = var_762_perm_0, x = var_761_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 14, 65, 64]> v_23_cast_fp16 = expand_dims(axes = v_23_axes_0, x = var_762_cast_fp16)[name = string("v_23_cast_fp16")];
            fp16 mul_11_y_0_to_fp16 = const()[name = string("mul_11_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_11_cast_fp16 = mul(x = q_23_cast_fp16, y = mul_11_y_0_to_fp16)[name = string("mul_11_cast_fp16")];
            bool matmul_11_transpose_y_0 = const()[name = string("matmul_11_transpose_y_0"), val = bool(true)];
            bool matmul_11_transpose_x_0 = const()[name = string("matmul_11_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_11_cast_fp16, y = k_23_cast_fp16)[name = string("matmul_11_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_11_cast_fp16 = add(x = matmul_11_cast_fp16, y = attn_mask_to_fp16)[name = string("add_11_cast_fp16")];
            int32 softmax_11_axis_0 = const()[name = string("softmax_11_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = add_11_cast_fp16)[name = string("softmax_11_cast_fp16")];
            bool a_23_transpose_x_0 = const()[name = string("a_23_transpose_x_0"), val = bool(false)];
            bool a_23_transpose_y_0 = const()[name = string("a_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_23_cast_fp16 = matmul(transpose_x = a_23_transpose_x_0, transpose_y = a_23_transpose_y_0, x = softmax_11_cast_fp16, y = v_23_cast_fp16)[name = string("a_23_cast_fp16")];
            tensor<int32, [1]> var_765_axes_0 = const()[name = string("op_765_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_765_cast_fp16 = squeeze(axes = var_765_axes_0, x = a_23_cast_fp16)[name = string("op_765_cast_fp16")];
            tensor<int32, [3]> var_766_perm_0 = const()[name = string("op_766_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_767 = const()[name = string("op_767"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_766_cast_fp16 = transpose(perm = var_766_perm_0, x = var_765_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [65, 896]> input_145_cast_fp16 = reshape(shape = var_767, x = var_766_cast_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_11_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119904256))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120707136))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_11_out_bias_to_fp16 = const()[name = string("trunk_layers_11_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120708992)))];
            tensor<fp16, [65, 896]> linear_70_cast_fp16 = linear(bias = trunk_layers_11_out_bias_to_fp16, weight = trunk_layers_11_out_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [65, 896]> input_147_cast_fp16 = add(x = input_141_cast_fp16, y = linear_70_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<int32, [1]> input_149_axes_0 = const()[name = string("input_149_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_11_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_11_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120710848)))];
            tensor<fp16, [896]> trunk_layers_11_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_11_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120712704)))];
            tensor<fp16, [65, 896]> input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = trunk_layers_11_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_11_mlp_ln_weight_to_fp16, x = input_147_cast_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_11_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120714560))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123925888))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_11_fc1_bias_to_fp16 = const()[name = string("trunk_layers_11_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123933120)))];
            tensor<fp16, [65, 3584]> linear_71_cast_fp16 = linear(bias = trunk_layers_11_fc1_bias_to_fp16, weight = trunk_layers_11_fc1_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_71_cast_fp16")];
            string input_151_mode_0 = const()[name = string("input_151_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = linear_71_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_11_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_11_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123940352))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127151680))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_11_fc2_bias_to_fp16 = const()[name = string("trunk_layers_11_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127153536)))];
            tensor<fp16, [65, 896]> linear_72_cast_fp16 = linear(bias = trunk_layers_11_fc2_bias_to_fp16, weight = trunk_layers_11_fc2_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<fp16, [65, 896]> input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_72_cast_fp16)[name = string("input_153_cast_fp16")];
            tensor<int32, [1]> input_155_axes_0 = const()[name = string("input_155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_12_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_12_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127155392)))];
            tensor<fp16, [896]> trunk_layers_12_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_12_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127157248)))];
            tensor<fp16, [65, 896]> input_155_cast_fp16 = layer_norm(axes = input_155_axes_0, beta = trunk_layers_12_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_12_attn_ln_weight_to_fp16, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_12_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127159104))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127961984))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_12_q_bias_to_fp16 = const()[name = string("trunk_layers_12_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127963840)))];
            tensor<fp16, [65, 896]> linear_73_cast_fp16 = linear(bias = trunk_layers_12_q_bias_to_fp16, weight = trunk_layers_12_q_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<int32, [3]> var_801 = const()[name = string("op_801"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_802_cast_fp16 = reshape(shape = var_801, x = linear_73_cast_fp16)[name = string("op_802_cast_fp16")];
            tensor<int32, [3]> var_803_perm_0 = const()[name = string("op_803_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_25_axes_0 = const()[name = string("q_25_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_803_cast_fp16 = transpose(perm = var_803_perm_0, x = var_802_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 14, 65, 64]> q_25_cast_fp16 = expand_dims(axes = q_25_axes_0, x = var_803_cast_fp16)[name = string("q_25_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_12_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127965696))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128768576))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_12_k_bias_to_fp16 = const()[name = string("trunk_layers_12_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128770432)))];
            tensor<fp16, [65, 896]> linear_74_cast_fp16 = linear(bias = trunk_layers_12_k_bias_to_fp16, weight = trunk_layers_12_k_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<int32, [3]> var_808 = const()[name = string("op_808"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_809_cast_fp16 = reshape(shape = var_808, x = linear_74_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<int32, [3]> var_810_perm_0 = const()[name = string("op_810_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_25_axes_0 = const()[name = string("k_25_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_810_cast_fp16 = transpose(perm = var_810_perm_0, x = var_809_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 14, 65, 64]> k_25_cast_fp16 = expand_dims(axes = k_25_axes_0, x = var_810_cast_fp16)[name = string("k_25_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_12_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128772288))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129575168))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_12_v_bias_to_fp16 = const()[name = string("trunk_layers_12_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129577024)))];
            tensor<fp16, [65, 896]> linear_75_cast_fp16 = linear(bias = trunk_layers_12_v_bias_to_fp16, weight = trunk_layers_12_v_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<int32, [3]> var_815 = const()[name = string("op_815"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_816_cast_fp16 = reshape(shape = var_815, x = linear_75_cast_fp16)[name = string("op_816_cast_fp16")];
            tensor<int32, [3]> var_817_perm_0 = const()[name = string("op_817_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_25_axes_0 = const()[name = string("v_25_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_817_cast_fp16 = transpose(perm = var_817_perm_0, x = var_816_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 14, 65, 64]> v_25_cast_fp16 = expand_dims(axes = v_25_axes_0, x = var_817_cast_fp16)[name = string("v_25_cast_fp16")];
            fp16 mul_12_y_0_to_fp16 = const()[name = string("mul_12_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_12_cast_fp16 = mul(x = q_25_cast_fp16, y = mul_12_y_0_to_fp16)[name = string("mul_12_cast_fp16")];
            bool matmul_12_transpose_y_0 = const()[name = string("matmul_12_transpose_y_0"), val = bool(true)];
            bool matmul_12_transpose_x_0 = const()[name = string("matmul_12_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = mul_12_cast_fp16, y = k_25_cast_fp16)[name = string("matmul_12_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_12_cast_fp16 = add(x = matmul_12_cast_fp16, y = attn_mask_to_fp16)[name = string("add_12_cast_fp16")];
            int32 softmax_12_axis_0 = const()[name = string("softmax_12_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = add_12_cast_fp16)[name = string("softmax_12_cast_fp16")];
            bool a_25_transpose_x_0 = const()[name = string("a_25_transpose_x_0"), val = bool(false)];
            bool a_25_transpose_y_0 = const()[name = string("a_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_25_cast_fp16 = matmul(transpose_x = a_25_transpose_x_0, transpose_y = a_25_transpose_y_0, x = softmax_12_cast_fp16, y = v_25_cast_fp16)[name = string("a_25_cast_fp16")];
            tensor<int32, [1]> var_820_axes_0 = const()[name = string("op_820_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_820_cast_fp16 = squeeze(axes = var_820_axes_0, x = a_25_cast_fp16)[name = string("op_820_cast_fp16")];
            tensor<int32, [3]> var_821_perm_0 = const()[name = string("op_821_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_822 = const()[name = string("op_822"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_821_cast_fp16 = transpose(perm = var_821_perm_0, x = var_820_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [65, 896]> input_157_cast_fp16 = reshape(shape = var_822, x = var_821_cast_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_12_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129578880))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130381760))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_12_out_bias_to_fp16 = const()[name = string("trunk_layers_12_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130383616)))];
            tensor<fp16, [65, 896]> linear_76_cast_fp16 = linear(bias = trunk_layers_12_out_bias_to_fp16, weight = trunk_layers_12_out_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [65, 896]> input_159_cast_fp16 = add(x = input_153_cast_fp16, y = linear_76_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<int32, [1]> input_161_axes_0 = const()[name = string("input_161_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_12_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_12_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130385472)))];
            tensor<fp16, [896]> trunk_layers_12_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_12_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130387328)))];
            tensor<fp16, [65, 896]> input_161_cast_fp16 = layer_norm(axes = input_161_axes_0, beta = trunk_layers_12_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_12_mlp_ln_weight_to_fp16, x = input_159_cast_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_12_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130389184))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133600512))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_12_fc1_bias_to_fp16 = const()[name = string("trunk_layers_12_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133607744)))];
            tensor<fp16, [65, 3584]> linear_77_cast_fp16 = linear(bias = trunk_layers_12_fc1_bias_to_fp16, weight = trunk_layers_12_fc1_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_77_cast_fp16")];
            string input_163_mode_0 = const()[name = string("input_163_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_163_cast_fp16 = gelu(mode = input_163_mode_0, x = linear_77_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_12_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_12_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133614976))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136826304))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_12_fc2_bias_to_fp16 = const()[name = string("trunk_layers_12_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136828160)))];
            tensor<fp16, [65, 896]> linear_78_cast_fp16 = linear(bias = trunk_layers_12_fc2_bias_to_fp16, weight = trunk_layers_12_fc2_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<fp16, [65, 896]> input_165_cast_fp16 = add(x = input_159_cast_fp16, y = linear_78_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<int32, [1]> input_167_axes_0 = const()[name = string("input_167_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_13_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_13_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136830016)))];
            tensor<fp16, [896]> trunk_layers_13_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_13_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136831872)))];
            tensor<fp16, [65, 896]> input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = trunk_layers_13_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_13_attn_ln_weight_to_fp16, x = input_165_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_13_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136833728))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137636608))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_13_q_bias_to_fp16 = const()[name = string("trunk_layers_13_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137638464)))];
            tensor<fp16, [65, 896]> linear_79_cast_fp16 = linear(bias = trunk_layers_13_q_bias_to_fp16, weight = trunk_layers_13_q_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [3]> var_856 = const()[name = string("op_856"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_857_cast_fp16 = reshape(shape = var_856, x = linear_79_cast_fp16)[name = string("op_857_cast_fp16")];
            tensor<int32, [3]> var_858_perm_0 = const()[name = string("op_858_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_27_axes_0 = const()[name = string("q_27_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_858_cast_fp16 = transpose(perm = var_858_perm_0, x = var_857_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 14, 65, 64]> q_27_cast_fp16 = expand_dims(axes = q_27_axes_0, x = var_858_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_13_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137640320))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138443200))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_13_k_bias_to_fp16 = const()[name = string("trunk_layers_13_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138445056)))];
            tensor<fp16, [65, 896]> linear_80_cast_fp16 = linear(bias = trunk_layers_13_k_bias_to_fp16, weight = trunk_layers_13_k_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<int32, [3]> var_863 = const()[name = string("op_863"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_864_cast_fp16 = reshape(shape = var_863, x = linear_80_cast_fp16)[name = string("op_864_cast_fp16")];
            tensor<int32, [3]> var_865_perm_0 = const()[name = string("op_865_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_27_axes_0 = const()[name = string("k_27_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_865_cast_fp16 = transpose(perm = var_865_perm_0, x = var_864_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 14, 65, 64]> k_27_cast_fp16 = expand_dims(axes = k_27_axes_0, x = var_865_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_13_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138446912))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139249792))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_13_v_bias_to_fp16 = const()[name = string("trunk_layers_13_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139251648)))];
            tensor<fp16, [65, 896]> linear_81_cast_fp16 = linear(bias = trunk_layers_13_v_bias_to_fp16, weight = trunk_layers_13_v_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<int32, [3]> var_870 = const()[name = string("op_870"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_871_cast_fp16 = reshape(shape = var_870, x = linear_81_cast_fp16)[name = string("op_871_cast_fp16")];
            tensor<int32, [3]> var_872_perm_0 = const()[name = string("op_872_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_27_axes_0 = const()[name = string("v_27_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_872_cast_fp16 = transpose(perm = var_872_perm_0, x = var_871_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 14, 65, 64]> v_27_cast_fp16 = expand_dims(axes = v_27_axes_0, x = var_872_cast_fp16)[name = string("v_27_cast_fp16")];
            fp16 mul_13_y_0_to_fp16 = const()[name = string("mul_13_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_13_cast_fp16 = mul(x = q_27_cast_fp16, y = mul_13_y_0_to_fp16)[name = string("mul_13_cast_fp16")];
            bool matmul_13_transpose_y_0 = const()[name = string("matmul_13_transpose_y_0"), val = bool(true)];
            bool matmul_13_transpose_x_0 = const()[name = string("matmul_13_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = mul_13_cast_fp16, y = k_27_cast_fp16)[name = string("matmul_13_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_13_cast_fp16 = add(x = matmul_13_cast_fp16, y = attn_mask_to_fp16)[name = string("add_13_cast_fp16")];
            int32 softmax_13_axis_0 = const()[name = string("softmax_13_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = add_13_cast_fp16)[name = string("softmax_13_cast_fp16")];
            bool a_27_transpose_x_0 = const()[name = string("a_27_transpose_x_0"), val = bool(false)];
            bool a_27_transpose_y_0 = const()[name = string("a_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_27_cast_fp16 = matmul(transpose_x = a_27_transpose_x_0, transpose_y = a_27_transpose_y_0, x = softmax_13_cast_fp16, y = v_27_cast_fp16)[name = string("a_27_cast_fp16")];
            tensor<int32, [1]> var_875_axes_0 = const()[name = string("op_875_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_875_cast_fp16 = squeeze(axes = var_875_axes_0, x = a_27_cast_fp16)[name = string("op_875_cast_fp16")];
            tensor<int32, [3]> var_876_perm_0 = const()[name = string("op_876_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_877 = const()[name = string("op_877"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_876_cast_fp16 = transpose(perm = var_876_perm_0, x = var_875_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [65, 896]> input_169_cast_fp16 = reshape(shape = var_877, x = var_876_cast_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_13_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139253504))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140056384))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_13_out_bias_to_fp16 = const()[name = string("trunk_layers_13_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140058240)))];
            tensor<fp16, [65, 896]> linear_82_cast_fp16 = linear(bias = trunk_layers_13_out_bias_to_fp16, weight = trunk_layers_13_out_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [65, 896]> input_171_cast_fp16 = add(x = input_165_cast_fp16, y = linear_82_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<int32, [1]> input_173_axes_0 = const()[name = string("input_173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_13_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_13_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140060096)))];
            tensor<fp16, [896]> trunk_layers_13_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_13_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140061952)))];
            tensor<fp16, [65, 896]> input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = trunk_layers_13_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_13_mlp_ln_weight_to_fp16, x = input_171_cast_fp16)[name = string("input_173_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_13_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140063808))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143275136))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_13_fc1_bias_to_fp16 = const()[name = string("trunk_layers_13_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143282368)))];
            tensor<fp16, [65, 3584]> linear_83_cast_fp16 = linear(bias = trunk_layers_13_fc1_bias_to_fp16, weight = trunk_layers_13_fc1_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_83_cast_fp16")];
            string input_175_mode_0 = const()[name = string("input_175_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_175_cast_fp16 = gelu(mode = input_175_mode_0, x = linear_83_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_13_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_13_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143289600))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146500928))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_13_fc2_bias_to_fp16 = const()[name = string("trunk_layers_13_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146502784)))];
            tensor<fp16, [65, 896]> linear_84_cast_fp16 = linear(bias = trunk_layers_13_fc2_bias_to_fp16, weight = trunk_layers_13_fc2_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<fp16, [65, 896]> input_177_cast_fp16 = add(x = input_171_cast_fp16, y = linear_84_cast_fp16)[name = string("input_177_cast_fp16")];
            tensor<int32, [1]> input_179_axes_0 = const()[name = string("input_179_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_14_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_14_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146504640)))];
            tensor<fp16, [896]> trunk_layers_14_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_14_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146506496)))];
            tensor<fp16, [65, 896]> input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = trunk_layers_14_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_14_attn_ln_weight_to_fp16, x = input_177_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_14_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146508352))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147311232))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_14_q_bias_to_fp16 = const()[name = string("trunk_layers_14_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147313088)))];
            tensor<fp16, [65, 896]> linear_85_cast_fp16 = linear(bias = trunk_layers_14_q_bias_to_fp16, weight = trunk_layers_14_q_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [3]> var_911 = const()[name = string("op_911"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_912_cast_fp16 = reshape(shape = var_911, x = linear_85_cast_fp16)[name = string("op_912_cast_fp16")];
            tensor<int32, [3]> var_913_perm_0 = const()[name = string("op_913_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_29_axes_0 = const()[name = string("q_29_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_913_cast_fp16 = transpose(perm = var_913_perm_0, x = var_912_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 14, 65, 64]> q_29_cast_fp16 = expand_dims(axes = q_29_axes_0, x = var_913_cast_fp16)[name = string("q_29_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_14_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147314944))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148117824))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_14_k_bias_to_fp16 = const()[name = string("trunk_layers_14_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148119680)))];
            tensor<fp16, [65, 896]> linear_86_cast_fp16 = linear(bias = trunk_layers_14_k_bias_to_fp16, weight = trunk_layers_14_k_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [3]> var_918 = const()[name = string("op_918"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_919_cast_fp16 = reshape(shape = var_918, x = linear_86_cast_fp16)[name = string("op_919_cast_fp16")];
            tensor<int32, [3]> var_920_perm_0 = const()[name = string("op_920_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_29_axes_0 = const()[name = string("k_29_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_920_cast_fp16 = transpose(perm = var_920_perm_0, x = var_919_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 14, 65, 64]> k_29_cast_fp16 = expand_dims(axes = k_29_axes_0, x = var_920_cast_fp16)[name = string("k_29_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_14_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148121536))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148924416))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_14_v_bias_to_fp16 = const()[name = string("trunk_layers_14_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148926272)))];
            tensor<fp16, [65, 896]> linear_87_cast_fp16 = linear(bias = trunk_layers_14_v_bias_to_fp16, weight = trunk_layers_14_v_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<int32, [3]> var_925 = const()[name = string("op_925"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_926_cast_fp16 = reshape(shape = var_925, x = linear_87_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [3]> var_927_perm_0 = const()[name = string("op_927_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_29_axes_0 = const()[name = string("v_29_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_927_cast_fp16 = transpose(perm = var_927_perm_0, x = var_926_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 14, 65, 64]> v_29_cast_fp16 = expand_dims(axes = v_29_axes_0, x = var_927_cast_fp16)[name = string("v_29_cast_fp16")];
            fp16 mul_14_y_0_to_fp16 = const()[name = string("mul_14_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_14_cast_fp16 = mul(x = q_29_cast_fp16, y = mul_14_y_0_to_fp16)[name = string("mul_14_cast_fp16")];
            bool matmul_14_transpose_y_0 = const()[name = string("matmul_14_transpose_y_0"), val = bool(true)];
            bool matmul_14_transpose_x_0 = const()[name = string("matmul_14_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = mul_14_cast_fp16, y = k_29_cast_fp16)[name = string("matmul_14_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_14_cast_fp16 = add(x = matmul_14_cast_fp16, y = attn_mask_to_fp16)[name = string("add_14_cast_fp16")];
            int32 softmax_14_axis_0 = const()[name = string("softmax_14_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = add_14_cast_fp16)[name = string("softmax_14_cast_fp16")];
            bool a_29_transpose_x_0 = const()[name = string("a_29_transpose_x_0"), val = bool(false)];
            bool a_29_transpose_y_0 = const()[name = string("a_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_29_cast_fp16 = matmul(transpose_x = a_29_transpose_x_0, transpose_y = a_29_transpose_y_0, x = softmax_14_cast_fp16, y = v_29_cast_fp16)[name = string("a_29_cast_fp16")];
            tensor<int32, [1]> var_930_axes_0 = const()[name = string("op_930_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_930_cast_fp16 = squeeze(axes = var_930_axes_0, x = a_29_cast_fp16)[name = string("op_930_cast_fp16")];
            tensor<int32, [3]> var_931_perm_0 = const()[name = string("op_931_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_932 = const()[name = string("op_932"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_931_cast_fp16 = transpose(perm = var_931_perm_0, x = var_930_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [65, 896]> input_181_cast_fp16 = reshape(shape = var_932, x = var_931_cast_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_14_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148928128))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149731008))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_14_out_bias_to_fp16 = const()[name = string("trunk_layers_14_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149732864)))];
            tensor<fp16, [65, 896]> linear_88_cast_fp16 = linear(bias = trunk_layers_14_out_bias_to_fp16, weight = trunk_layers_14_out_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [65, 896]> input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_88_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<int32, [1]> input_185_axes_0 = const()[name = string("input_185_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_14_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_14_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149734720)))];
            tensor<fp16, [896]> trunk_layers_14_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_14_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149736576)))];
            tensor<fp16, [65, 896]> input_185_cast_fp16 = layer_norm(axes = input_185_axes_0, beta = trunk_layers_14_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_14_mlp_ln_weight_to_fp16, x = input_183_cast_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_14_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149738432))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152949760))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_14_fc1_bias_to_fp16 = const()[name = string("trunk_layers_14_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152956992)))];
            tensor<fp16, [65, 3584]> linear_89_cast_fp16 = linear(bias = trunk_layers_14_fc1_bias_to_fp16, weight = trunk_layers_14_fc1_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_89_cast_fp16")];
            string input_187_mode_0 = const()[name = string("input_187_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_187_cast_fp16 = gelu(mode = input_187_mode_0, x = linear_89_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_14_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_14_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152964224))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156175552))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_14_fc2_bias_to_fp16 = const()[name = string("trunk_layers_14_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156177408)))];
            tensor<fp16, [65, 896]> linear_90_cast_fp16 = linear(bias = trunk_layers_14_fc2_bias_to_fp16, weight = trunk_layers_14_fc2_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [65, 896]> input_189_cast_fp16 = add(x = input_183_cast_fp16, y = linear_90_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<int32, [1]> input_191_axes_0 = const()[name = string("input_191_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_15_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_15_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156179264)))];
            tensor<fp16, [896]> trunk_layers_15_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_15_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156181120)))];
            tensor<fp16, [65, 896]> input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = trunk_layers_15_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_15_attn_ln_weight_to_fp16, x = input_189_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_15_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156182976))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156985856))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_15_q_bias_to_fp16 = const()[name = string("trunk_layers_15_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156987712)))];
            tensor<fp16, [65, 896]> linear_91_cast_fp16 = linear(bias = trunk_layers_15_q_bias_to_fp16, weight = trunk_layers_15_q_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [3]> var_966 = const()[name = string("op_966"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_967_cast_fp16 = reshape(shape = var_966, x = linear_91_cast_fp16)[name = string("op_967_cast_fp16")];
            tensor<int32, [3]> var_968_perm_0 = const()[name = string("op_968_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_31_axes_0 = const()[name = string("q_31_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_968_cast_fp16 = transpose(perm = var_968_perm_0, x = var_967_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 14, 65, 64]> q_31_cast_fp16 = expand_dims(axes = q_31_axes_0, x = var_968_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_15_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156989568))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157792448))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_15_k_bias_to_fp16 = const()[name = string("trunk_layers_15_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157794304)))];
            tensor<fp16, [65, 896]> linear_92_cast_fp16 = linear(bias = trunk_layers_15_k_bias_to_fp16, weight = trunk_layers_15_k_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [3]> var_973 = const()[name = string("op_973"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_974_cast_fp16 = reshape(shape = var_973, x = linear_92_cast_fp16)[name = string("op_974_cast_fp16")];
            tensor<int32, [3]> var_975_perm_0 = const()[name = string("op_975_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_31_axes_0 = const()[name = string("k_31_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_975_cast_fp16 = transpose(perm = var_975_perm_0, x = var_974_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 14, 65, 64]> k_31_cast_fp16 = expand_dims(axes = k_31_axes_0, x = var_975_cast_fp16)[name = string("k_31_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_15_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157796160))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158599040))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_15_v_bias_to_fp16 = const()[name = string("trunk_layers_15_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158600896)))];
            tensor<fp16, [65, 896]> linear_93_cast_fp16 = linear(bias = trunk_layers_15_v_bias_to_fp16, weight = trunk_layers_15_v_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [3]> var_980 = const()[name = string("op_980"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_981_cast_fp16 = reshape(shape = var_980, x = linear_93_cast_fp16)[name = string("op_981_cast_fp16")];
            tensor<int32, [3]> var_982_perm_0 = const()[name = string("op_982_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_31_axes_0 = const()[name = string("v_31_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_982_cast_fp16 = transpose(perm = var_982_perm_0, x = var_981_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 14, 65, 64]> v_31_cast_fp16 = expand_dims(axes = v_31_axes_0, x = var_982_cast_fp16)[name = string("v_31_cast_fp16")];
            fp16 mul_15_y_0_to_fp16 = const()[name = string("mul_15_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_15_cast_fp16 = mul(x = q_31_cast_fp16, y = mul_15_y_0_to_fp16)[name = string("mul_15_cast_fp16")];
            bool matmul_15_transpose_y_0 = const()[name = string("matmul_15_transpose_y_0"), val = bool(true)];
            bool matmul_15_transpose_x_0 = const()[name = string("matmul_15_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = mul_15_cast_fp16, y = k_31_cast_fp16)[name = string("matmul_15_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_15_cast_fp16 = add(x = matmul_15_cast_fp16, y = attn_mask_to_fp16)[name = string("add_15_cast_fp16")];
            int32 softmax_15_axis_0 = const()[name = string("softmax_15_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = add_15_cast_fp16)[name = string("softmax_15_cast_fp16")];
            bool a_31_transpose_x_0 = const()[name = string("a_31_transpose_x_0"), val = bool(false)];
            bool a_31_transpose_y_0 = const()[name = string("a_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_31_cast_fp16 = matmul(transpose_x = a_31_transpose_x_0, transpose_y = a_31_transpose_y_0, x = softmax_15_cast_fp16, y = v_31_cast_fp16)[name = string("a_31_cast_fp16")];
            tensor<int32, [1]> var_985_axes_0 = const()[name = string("op_985_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_985_cast_fp16 = squeeze(axes = var_985_axes_0, x = a_31_cast_fp16)[name = string("op_985_cast_fp16")];
            tensor<int32, [3]> var_986_perm_0 = const()[name = string("op_986_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_987 = const()[name = string("op_987"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_986_cast_fp16 = transpose(perm = var_986_perm_0, x = var_985_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [65, 896]> input_193_cast_fp16 = reshape(shape = var_987, x = var_986_cast_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_15_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158602752))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159405632))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_15_out_bias_to_fp16 = const()[name = string("trunk_layers_15_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159407488)))];
            tensor<fp16, [65, 896]> linear_94_cast_fp16 = linear(bias = trunk_layers_15_out_bias_to_fp16, weight = trunk_layers_15_out_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [65, 896]> input_195_cast_fp16 = add(x = input_189_cast_fp16, y = linear_94_cast_fp16)[name = string("input_195_cast_fp16")];
            tensor<int32, [1]> input_197_axes_0 = const()[name = string("input_197_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_15_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_15_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159409344)))];
            tensor<fp16, [896]> trunk_layers_15_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_15_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159411200)))];
            tensor<fp16, [65, 896]> input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = trunk_layers_15_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_15_mlp_ln_weight_to_fp16, x = input_195_cast_fp16)[name = string("input_197_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_15_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159413056))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162624384))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_15_fc1_bias_to_fp16 = const()[name = string("trunk_layers_15_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162631616)))];
            tensor<fp16, [65, 3584]> linear_95_cast_fp16 = linear(bias = trunk_layers_15_fc1_bias_to_fp16, weight = trunk_layers_15_fc1_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_95_cast_fp16")];
            string input_199_mode_0 = const()[name = string("input_199_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = linear_95_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_15_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_15_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162638848))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165850176))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_15_fc2_bias_to_fp16 = const()[name = string("trunk_layers_15_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165852032)))];
            tensor<fp16, [65, 896]> linear_96_cast_fp16 = linear(bias = trunk_layers_15_fc2_bias_to_fp16, weight = trunk_layers_15_fc2_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [65, 896]> input_201_cast_fp16 = add(x = input_195_cast_fp16, y = linear_96_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<int32, [1]> input_203_axes_0 = const()[name = string("input_203_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_16_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_16_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165853888)))];
            tensor<fp16, [896]> trunk_layers_16_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_16_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165855744)))];
            tensor<fp16, [65, 896]> input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = trunk_layers_16_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_16_attn_ln_weight_to_fp16, x = input_201_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_16_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165857600))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166660480))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_16_q_bias_to_fp16 = const()[name = string("trunk_layers_16_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166662336)))];
            tensor<fp16, [65, 896]> linear_97_cast_fp16 = linear(bias = trunk_layers_16_q_bias_to_fp16, weight = trunk_layers_16_q_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<int32, [3]> var_1021 = const()[name = string("op_1021"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_1022_cast_fp16 = reshape(shape = var_1021, x = linear_97_cast_fp16)[name = string("op_1022_cast_fp16")];
            tensor<int32, [3]> var_1023_perm_0 = const()[name = string("op_1023_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_33_axes_0 = const()[name = string("q_33_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_1023_cast_fp16 = transpose(perm = var_1023_perm_0, x = var_1022_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 14, 65, 64]> q_33_cast_fp16 = expand_dims(axes = q_33_axes_0, x = var_1023_cast_fp16)[name = string("q_33_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_16_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166664192))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167467072))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_16_k_bias_to_fp16 = const()[name = string("trunk_layers_16_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167468928)))];
            tensor<fp16, [65, 896]> linear_98_cast_fp16 = linear(bias = trunk_layers_16_k_bias_to_fp16, weight = trunk_layers_16_k_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [3]> var_1028 = const()[name = string("op_1028"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_1029_cast_fp16 = reshape(shape = var_1028, x = linear_98_cast_fp16)[name = string("op_1029_cast_fp16")];
            tensor<int32, [3]> var_1030_perm_0 = const()[name = string("op_1030_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_33_axes_0 = const()[name = string("k_33_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_1030_cast_fp16 = transpose(perm = var_1030_perm_0, x = var_1029_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 14, 65, 64]> k_33_cast_fp16 = expand_dims(axes = k_33_axes_0, x = var_1030_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_16_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167470784))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168273664))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_16_v_bias_to_fp16 = const()[name = string("trunk_layers_16_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168275520)))];
            tensor<fp16, [65, 896]> linear_99_cast_fp16 = linear(bias = trunk_layers_16_v_bias_to_fp16, weight = trunk_layers_16_v_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [3]> var_1035 = const()[name = string("op_1035"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_1036_cast_fp16 = reshape(shape = var_1035, x = linear_99_cast_fp16)[name = string("op_1036_cast_fp16")];
            tensor<int32, [3]> var_1037_perm_0 = const()[name = string("op_1037_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_33_axes_0 = const()[name = string("v_33_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_1037_cast_fp16 = transpose(perm = var_1037_perm_0, x = var_1036_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 14, 65, 64]> v_33_cast_fp16 = expand_dims(axes = v_33_axes_0, x = var_1037_cast_fp16)[name = string("v_33_cast_fp16")];
            fp16 mul_16_y_0_to_fp16 = const()[name = string("mul_16_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_16_cast_fp16 = mul(x = q_33_cast_fp16, y = mul_16_y_0_to_fp16)[name = string("mul_16_cast_fp16")];
            bool matmul_16_transpose_y_0 = const()[name = string("matmul_16_transpose_y_0"), val = bool(true)];
            bool matmul_16_transpose_x_0 = const()[name = string("matmul_16_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = mul_16_cast_fp16, y = k_33_cast_fp16)[name = string("matmul_16_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_16_cast_fp16 = add(x = matmul_16_cast_fp16, y = attn_mask_to_fp16)[name = string("add_16_cast_fp16")];
            int32 softmax_16_axis_0 = const()[name = string("softmax_16_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_16_cast_fp16 = softmax(axis = softmax_16_axis_0, x = add_16_cast_fp16)[name = string("softmax_16_cast_fp16")];
            bool a_33_transpose_x_0 = const()[name = string("a_33_transpose_x_0"), val = bool(false)];
            bool a_33_transpose_y_0 = const()[name = string("a_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_33_cast_fp16 = matmul(transpose_x = a_33_transpose_x_0, transpose_y = a_33_transpose_y_0, x = softmax_16_cast_fp16, y = v_33_cast_fp16)[name = string("a_33_cast_fp16")];
            tensor<int32, [1]> var_1040_axes_0 = const()[name = string("op_1040_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_1040_cast_fp16 = squeeze(axes = var_1040_axes_0, x = a_33_cast_fp16)[name = string("op_1040_cast_fp16")];
            tensor<int32, [3]> var_1041_perm_0 = const()[name = string("op_1041_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_1042 = const()[name = string("op_1042"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_1041_cast_fp16 = transpose(perm = var_1041_perm_0, x = var_1040_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [65, 896]> input_205_cast_fp16 = reshape(shape = var_1042, x = var_1041_cast_fp16)[name = string("input_205_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_16_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168277376))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169080256))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_16_out_bias_to_fp16 = const()[name = string("trunk_layers_16_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169082112)))];
            tensor<fp16, [65, 896]> linear_100_cast_fp16 = linear(bias = trunk_layers_16_out_bias_to_fp16, weight = trunk_layers_16_out_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<fp16, [65, 896]> input_207_cast_fp16 = add(x = input_201_cast_fp16, y = linear_100_cast_fp16)[name = string("input_207_cast_fp16")];
            tensor<int32, [1]> input_209_axes_0 = const()[name = string("input_209_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_16_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_16_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169083968)))];
            tensor<fp16, [896]> trunk_layers_16_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_16_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169085824)))];
            tensor<fp16, [65, 896]> input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = trunk_layers_16_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_16_mlp_ln_weight_to_fp16, x = input_207_cast_fp16)[name = string("input_209_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_16_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169087680))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172299008))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_16_fc1_bias_to_fp16 = const()[name = string("trunk_layers_16_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172306240)))];
            tensor<fp16, [65, 3584]> linear_101_cast_fp16 = linear(bias = trunk_layers_16_fc1_bias_to_fp16, weight = trunk_layers_16_fc1_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_101_cast_fp16")];
            string input_211_mode_0 = const()[name = string("input_211_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_211_cast_fp16 = gelu(mode = input_211_mode_0, x = linear_101_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_16_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_16_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172313472))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175524800))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_16_fc2_bias_to_fp16 = const()[name = string("trunk_layers_16_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175526656)))];
            tensor<fp16, [65, 896]> linear_102_cast_fp16 = linear(bias = trunk_layers_16_fc2_bias_to_fp16, weight = trunk_layers_16_fc2_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [65, 896]> input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_102_cast_fp16)[name = string("input_213_cast_fp16")];
            tensor<int32, [1]> input_215_axes_0 = const()[name = string("input_215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_17_attn_ln_weight_to_fp16 = const()[name = string("trunk_layers_17_attn_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175528512)))];
            tensor<fp16, [896]> trunk_layers_17_attn_ln_bias_to_fp16 = const()[name = string("trunk_layers_17_attn_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175530368)))];
            tensor<fp16, [65, 896]> input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = trunk_layers_17_attn_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_17_attn_ln_weight_to_fp16, x = input_213_cast_fp16)[name = string("input_215_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_17_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175532224))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176335104))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_17_q_bias_to_fp16 = const()[name = string("trunk_layers_17_q_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176336960)))];
            tensor<fp16, [65, 896]> linear_103_cast_fp16 = linear(bias = trunk_layers_17_q_bias_to_fp16, weight = trunk_layers_17_q_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<int32, [3]> var_1076 = const()[name = string("op_1076"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_1077_cast_fp16 = reshape(shape = var_1076, x = linear_103_cast_fp16)[name = string("op_1077_cast_fp16")];
            tensor<int32, [3]> var_1078_perm_0 = const()[name = string("op_1078_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> q_35_axes_0 = const()[name = string("q_35_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_1078_cast_fp16 = transpose(perm = var_1078_perm_0, x = var_1077_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 14, 65, 64]> q_35_cast_fp16 = expand_dims(axes = q_35_axes_0, x = var_1078_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_17_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176338816))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177141696))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_17_k_bias_to_fp16 = const()[name = string("trunk_layers_17_k_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177143552)))];
            tensor<fp16, [65, 896]> linear_104_cast_fp16 = linear(bias = trunk_layers_17_k_bias_to_fp16, weight = trunk_layers_17_k_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<int32, [3]> var_1083 = const()[name = string("op_1083"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_1084_cast_fp16 = reshape(shape = var_1083, x = linear_104_cast_fp16)[name = string("op_1084_cast_fp16")];
            tensor<int32, [3]> var_1085_perm_0 = const()[name = string("op_1085_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> k_35_axes_0 = const()[name = string("k_35_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_1085_cast_fp16 = transpose(perm = var_1085_perm_0, x = var_1084_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 14, 65, 64]> k_35_cast_fp16 = expand_dims(axes = k_35_axes_0, x = var_1085_cast_fp16)[name = string("k_35_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_17_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177145408))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177948288))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_17_v_bias_to_fp16 = const()[name = string("trunk_layers_17_v_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177950144)))];
            tensor<fp16, [65, 896]> linear_105_cast_fp16 = linear(bias = trunk_layers_17_v_bias_to_fp16, weight = trunk_layers_17_v_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [3]> var_1090 = const()[name = string("op_1090"), val = tensor<int32, [3]>([65, 14, 64])];
            tensor<fp16, [65, 14, 64]> var_1091_cast_fp16 = reshape(shape = var_1090, x = linear_105_cast_fp16)[name = string("op_1091_cast_fp16")];
            tensor<int32, [3]> var_1092_perm_0 = const()[name = string("op_1092_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> v_35_axes_0 = const()[name = string("v_35_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_1092_cast_fp16 = transpose(perm = var_1092_perm_0, x = var_1091_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 14, 65, 64]> v_35_cast_fp16 = expand_dims(axes = v_35_axes_0, x = var_1092_cast_fp16)[name = string("v_35_cast_fp16")];
            fp16 mul_17_y_0_to_fp16 = const()[name = string("mul_17_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 65, 64]> mul_17_cast_fp16 = mul(x = q_35_cast_fp16, y = mul_17_y_0_to_fp16)[name = string("mul_17_cast_fp16")];
            bool matmul_17_transpose_y_0 = const()[name = string("matmul_17_transpose_y_0"), val = bool(true)];
            bool matmul_17_transpose_x_0 = const()[name = string("matmul_17_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 65]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = mul_17_cast_fp16, y = k_35_cast_fp16)[name = string("matmul_17_cast_fp16")];
            tensor<fp16, [1, 14, 65, 65]> add_17_cast_fp16 = add(x = matmul_17_cast_fp16, y = attn_mask_to_fp16)[name = string("add_17_cast_fp16")];
            int32 softmax_17_axis_0 = const()[name = string("softmax_17_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 65, 65]> softmax_17_cast_fp16 = softmax(axis = softmax_17_axis_0, x = add_17_cast_fp16)[name = string("softmax_17_cast_fp16")];
            bool a_transpose_x_0 = const()[name = string("a_transpose_x_0"), val = bool(false)];
            bool a_transpose_y_0 = const()[name = string("a_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 65, 64]> a_cast_fp16 = matmul(transpose_x = a_transpose_x_0, transpose_y = a_transpose_y_0, x = softmax_17_cast_fp16, y = v_35_cast_fp16)[name = string("a_cast_fp16")];
            tensor<int32, [1]> var_1095_axes_0 = const()[name = string("op_1095_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [14, 65, 64]> var_1095_cast_fp16 = squeeze(axes = var_1095_axes_0, x = a_cast_fp16)[name = string("op_1095_cast_fp16")];
            tensor<int32, [3]> var_1096_perm_0 = const()[name = string("op_1096_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_1097 = const()[name = string("op_1097"), val = tensor<int32, [2]>([65, 896])];
            tensor<fp16, [65, 14, 64]> var_1096_cast_fp16 = transpose(perm = var_1096_perm_0, x = var_1095_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [65, 896]> input_217_cast_fp16 = reshape(shape = var_1097, x = var_1096_cast_fp16)[name = string("input_217_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_layers_17_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177952000))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178754880))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_17_out_bias_to_fp16 = const()[name = string("trunk_layers_17_out_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178756736)))];
            tensor<fp16, [65, 896]> linear_106_cast_fp16 = linear(bias = trunk_layers_17_out_bias_to_fp16, weight = trunk_layers_17_out_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<fp16, [65, 896]> input_219_cast_fp16 = add(x = input_213_cast_fp16, y = linear_106_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<int32, [1]> input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_layers_17_mlp_ln_weight_to_fp16 = const()[name = string("trunk_layers_17_mlp_ln_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178758592)))];
            tensor<fp16, [896]> trunk_layers_17_mlp_ln_bias_to_fp16 = const()[name = string("trunk_layers_17_mlp_ln_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178760448)))];
            tensor<fp16, [65, 896]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = trunk_layers_17_mlp_ln_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_layers_17_mlp_ln_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [3584, 896]> trunk_layers_17_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [3584, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178762304))), scale = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181973632))), zero_point = tensor<int8, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17501376)))];
            tensor<fp16, [3584]> trunk_layers_17_fc1_bias_to_fp16 = const()[name = string("trunk_layers_17_fc1_bias_to_fp16"), val = tensor<fp16, [3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181980864)))];
            tensor<fp16, [65, 3584]> linear_107_cast_fp16 = linear(bias = trunk_layers_17_fc1_bias_to_fp16, weight = trunk_layers_17_fc1_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_107_cast_fp16")];
            string input_223_mode_0 = const()[name = string("input_223_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 3584]> input_223_cast_fp16 = gelu(mode = input_223_mode_0, x = linear_107_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [896, 3584]> trunk_layers_17_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_layers_17_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 3584]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181988096))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185199424))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_layers_17_fc2_bias_to_fp16 = const()[name = string("trunk_layers_17_fc2_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185201280)))];
            tensor<fp16, [65, 896]> linear_108_cast_fp16 = linear(bias = trunk_layers_17_fc2_bias_to_fp16, weight = trunk_layers_17_fc2_weight_to_fp16_quantized, x = input_223_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [65, 896]> input_225_cast_fp16 = add(x = input_219_cast_fp16, y = linear_108_cast_fp16)[name = string("input_225_cast_fp16")];
            tensor<int32, [1]> input_227_axes_0 = const()[name = string("input_227_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [896]> trunk_ln_post_weight_to_fp16 = const()[name = string("trunk_ln_post_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185203136)))];
            tensor<fp16, [896]> trunk_ln_post_bias_to_fp16 = const()[name = string("trunk_ln_post_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185204992)))];
            tensor<fp16, [65, 896]> input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = trunk_ln_post_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_ln_post_weight_to_fp16, x = input_225_cast_fp16)[name = string("input_227_cast_fp16")];
            tensor<fp16, [896, 896]> trunk_proj1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_proj1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185206848))), scale = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186009728))), zero_point = tensor<int8, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11039872)))];
            tensor<fp16, [896]> trunk_proj1_bias_to_fp16 = const()[name = string("trunk_proj1_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186011584)))];
            tensor<fp16, [65, 896]> linear_109_cast_fp16 = linear(bias = trunk_proj1_bias_to_fp16, weight = trunk_proj1_weight_to_fp16_quantized, x = input_227_cast_fp16)[name = string("linear_109_cast_fp16")];
            string input_229_mode_0 = const()[name = string("input_229_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 896]> input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = linear_109_cast_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [1024, 896]> trunk_proj2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_proj2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186013440))), scale = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186932096))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186931008)))];
            tensor<fp16, [1024]> trunk_proj2_bias_to_fp16 = const()[name = string("trunk_proj2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186934208)))];
            tensor<fp16, [65, 1024]> linear_110_cast_fp16 = linear(bias = trunk_proj2_bias_to_fp16, weight = trunk_proj2_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<int32, [1]> input_233_axes_0 = const()[name = string("input_233_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> trunk_tower_0_norm_weight_to_fp16 = const()[name = string("trunk_tower_0_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186936320)))];
            tensor<fp16, [1024]> trunk_tower_0_norm_bias_to_fp16 = const()[name = string("trunk_tower_0_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186938432)))];
            tensor<fp16, [65, 1024]> input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = trunk_tower_0_norm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_tower_0_norm_weight_to_fp16, x = linear_110_cast_fp16)[name = string("input_233_cast_fp16")];
            tensor<fp16, [4096, 1024]> trunk_tower_0_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_0_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186940544))), scale = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191139072))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191134912)))];
            tensor<fp16, [4096]> trunk_tower_0_fc1_bias_to_fp16 = const()[name = string("trunk_tower_0_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191147328)))];
            tensor<fp16, [65, 4096]> linear_111_cast_fp16 = linear(bias = trunk_tower_0_fc1_bias_to_fp16, weight = trunk_tower_0_fc1_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_111_cast_fp16")];
            string input_235_mode_0 = const()[name = string("input_235_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 4096]> input_235_cast_fp16 = gelu(mode = input_235_mode_0, x = linear_111_cast_fp16)[name = string("input_235_cast_fp16")];
            tensor<fp16, [1024, 4096]> trunk_tower_0_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_0_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191155584))), scale = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195349952))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186931008)))];
            tensor<fp16, [1024]> trunk_tower_0_fc2_bias_to_fp16 = const()[name = string("trunk_tower_0_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195352064)))];
            tensor<fp16, [65, 1024]> linear_112_cast_fp16 = linear(bias = trunk_tower_0_fc2_bias_to_fp16, weight = trunk_tower_0_fc2_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<fp16, [65, 1024]> input_237_cast_fp16 = add(x = linear_110_cast_fp16, y = linear_112_cast_fp16)[name = string("input_237_cast_fp16")];
            tensor<int32, [1]> input_239_axes_0 = const()[name = string("input_239_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> trunk_tower_1_norm_weight_to_fp16 = const()[name = string("trunk_tower_1_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195354176)))];
            tensor<fp16, [1024]> trunk_tower_1_norm_bias_to_fp16 = const()[name = string("trunk_tower_1_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195356288)))];
            tensor<fp16, [65, 1024]> input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = trunk_tower_1_norm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_tower_1_norm_weight_to_fp16, x = input_237_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [4096, 1024]> trunk_tower_1_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_1_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195358400))), scale = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199552768))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191134912)))];
            tensor<fp16, [4096]> trunk_tower_1_fc1_bias_to_fp16 = const()[name = string("trunk_tower_1_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199561024)))];
            tensor<fp16, [65, 4096]> linear_113_cast_fp16 = linear(bias = trunk_tower_1_fc1_bias_to_fp16, weight = trunk_tower_1_fc1_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_113_cast_fp16")];
            string input_241_mode_0 = const()[name = string("input_241_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 4096]> input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_113_cast_fp16)[name = string("input_241_cast_fp16")];
            tensor<fp16, [1024, 4096]> trunk_tower_1_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_1_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199569280))), scale = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203763648))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186931008)))];
            tensor<fp16, [1024]> trunk_tower_1_fc2_bias_to_fp16 = const()[name = string("trunk_tower_1_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203765760)))];
            tensor<fp16, [65, 1024]> linear_114_cast_fp16 = linear(bias = trunk_tower_1_fc2_bias_to_fp16, weight = trunk_tower_1_fc2_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<fp16, [65, 1024]> input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_114_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<int32, [1]> input_245_axes_0 = const()[name = string("input_245_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> trunk_tower_2_norm_weight_to_fp16 = const()[name = string("trunk_tower_2_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203767872)))];
            tensor<fp16, [1024]> trunk_tower_2_norm_bias_to_fp16 = const()[name = string("trunk_tower_2_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203769984)))];
            tensor<fp16, [65, 1024]> input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = trunk_tower_2_norm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_tower_2_norm_weight_to_fp16, x = input_243_cast_fp16)[name = string("input_245_cast_fp16")];
            tensor<fp16, [4096, 1024]> trunk_tower_2_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_2_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203772096))), scale = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207966464))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191134912)))];
            tensor<fp16, [4096]> trunk_tower_2_fc1_bias_to_fp16 = const()[name = string("trunk_tower_2_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207974720)))];
            tensor<fp16, [65, 4096]> linear_115_cast_fp16 = linear(bias = trunk_tower_2_fc1_bias_to_fp16, weight = trunk_tower_2_fc1_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_115_cast_fp16")];
            string input_247_mode_0 = const()[name = string("input_247_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 4096]> input_247_cast_fp16 = gelu(mode = input_247_mode_0, x = linear_115_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<fp16, [1024, 4096]> trunk_tower_2_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_2_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207982976))), scale = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212177344))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186931008)))];
            tensor<fp16, [1024]> trunk_tower_2_fc2_bias_to_fp16 = const()[name = string("trunk_tower_2_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212179456)))];
            tensor<fp16, [65, 1024]> linear_116_cast_fp16 = linear(bias = trunk_tower_2_fc2_bias_to_fp16, weight = trunk_tower_2_fc2_weight_to_fp16_quantized, x = input_247_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [65, 1024]> input_249_cast_fp16 = add(x = input_243_cast_fp16, y = linear_116_cast_fp16)[name = string("input_249_cast_fp16")];
            tensor<int32, [1]> input_251_axes_0 = const()[name = string("input_251_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> trunk_tower_3_norm_weight_to_fp16 = const()[name = string("trunk_tower_3_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212181568)))];
            tensor<fp16, [1024]> trunk_tower_3_norm_bias_to_fp16 = const()[name = string("trunk_tower_3_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212183680)))];
            tensor<fp16, [65, 1024]> input_251_cast_fp16 = layer_norm(axes = input_251_axes_0, beta = trunk_tower_3_norm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_tower_3_norm_weight_to_fp16, x = input_249_cast_fp16)[name = string("input_251_cast_fp16")];
            tensor<fp16, [4096, 1024]> trunk_tower_3_fc1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_3_fc1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212185792))), scale = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216380160))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191134912)))];
            tensor<fp16, [4096]> trunk_tower_3_fc1_bias_to_fp16 = const()[name = string("trunk_tower_3_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216388416)))];
            tensor<fp16, [65, 4096]> linear_117_cast_fp16 = linear(bias = trunk_tower_3_fc1_bias_to_fp16, weight = trunk_tower_3_fc1_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_117_cast_fp16")];
            string input_253_mode_0 = const()[name = string("input_253_mode_0"), val = string("EXACT")];
            tensor<fp16, [65, 4096]> input_253_cast_fp16 = gelu(mode = input_253_mode_0, x = linear_117_cast_fp16)[name = string("input_253_cast_fp16")];
            tensor<fp16, [1024, 4096]> trunk_tower_3_fc2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("trunk_tower_3_fc2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216396672))), scale = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220591040))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186931008)))];
            tensor<fp16, [1024]> trunk_tower_3_fc2_bias_to_fp16 = const()[name = string("trunk_tower_3_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220593152)))];
            tensor<fp16, [65, 1024]> linear_118_cast_fp16 = linear(bias = trunk_tower_3_fc2_bias_to_fp16, weight = trunk_tower_3_fc2_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [65, 1024]> input_cast_fp16 = add(x = input_249_cast_fp16, y = linear_118_cast_fp16)[name = string("input_cast_fp16")];
            tensor<int32, [1]> var_1189_axes_0 = const()[name = string("op_1189_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> trunk_final_norm_weight_to_fp16 = const()[name = string("trunk_final_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220595264)))];
            tensor<fp16, [1024]> trunk_final_norm_bias_to_fp16 = const()[name = string("trunk_final_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220597376)))];
            tensor<fp16, [65, 1024]> hidden = layer_norm(axes = var_1189_axes_0, beta = trunk_final_norm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = trunk_final_norm_weight_to_fp16, x = input_cast_fp16)[name = string("op_1189_cast_fp16")];
        } -> (hidden);
}