diff --git "a/vae/modelopt-component-manifest.json" "b/vae/modelopt-component-manifest.json" new file mode 100644--- /dev/null +++ "b/vae/modelopt-component-manifest.json" @@ -0,0 +1,3063 @@ +{ + "component": "video_vae", + "files": { + "config.json": { + "bytes": 2011, + "sha256": "78f67deec3d63aae807f2bfe7154bc1e26f6372cb20b63265fcbae1b62bb5745" + }, + "diffusion_pytorch_model-00001-of-00003.safetensors": { + "bytes": 735686172, + "sha256": "45c80b94a721ec95956775054b3abbf3ae7c61bae695136c4a6e32be5884d3e1" + }, + "diffusion_pytorch_model-00002-of-00003.safetensors": { + "bytes": 976666252, + "sha256": "c8ffb841645a7ac8a30be90ed957cb4d61041c4efb8b95509dc8fb18abe8af89" + }, + "diffusion_pytorch_model-00003-of-00003.safetensors": { + "bytes": 398539336, + "sha256": "c05d6ac4b1a33de372799d708531da6320f6a3ce6d1ce6d895e770988e004a39" + }, + "diffusion_pytorch_model.safetensors.index.json": { + "bytes": 126885, + "sha256": "79a3aef4ddc871bd5c4d72406a9f1fa9ff2c12df0cfa83fbcbcd98a24d04fd4b" + } + }, + "modelopt_commit": "6a4b3f147e14a6fec690fedbced8df402344085d", + "numerical_format": { + "activation": "a4", + "block_scale": "fp8_e4m3", + "block_size": 16, + "output": "bf16", + "tensor_scale": "fp32", + "values": "e2m1", + "weight": "w4" + }, + "source": "/workspace/models/FastVideo-FastH3-8-Step-V2/vae", + "target_weights": 216, + "weights": { + "decoder.transformer_blocks.0.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 97.90484619140625, + "block_scale_min": 30.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.11448065936565399, + "weight_saturation_rate": 0.06715750694274902, + "weight_tensor_scale": 4.258953049429692e-05 + }, + "decoder.transformer_blocks.0.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 137.61119079589844, + "block_scale_min": 24.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.09913107007741928, + "weight_saturation_rate": 0.06422543525695801, + "weight_tensor_scale": 3.68791188520845e-05 + }, + "decoder.transformer_blocks.0.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 62.39039611816406, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.18064333498477936, + "weight_saturation_rate": 0.059830665588378906, + "weight_tensor_scale": 6.720362580381334e-05 + }, + "decoder.transformer_blocks.0.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 87.70352172851562, + "block_scale_min": 26.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.14241331815719604, + "weight_saturation_rate": 0.06371736526489258, + "weight_tensor_scale": 5.2981147746322677e-05 + }, + "decoder.transformer_blocks.0.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 25.457622528076172, + "block_scale_min": 6.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.3599620461463928, + "weight_saturation_rate": 0.06585067510604858, + "weight_tensor_scale": 0.00013391445099841803 + }, + "decoder.transformer_blocks.0.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 89.32765197753906, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.11548785120248795, + "weight_saturation_rate": 0.06434088945388794, + "weight_tensor_scale": 4.296423139749095e-05 + }, + "decoder.transformer_blocks.1.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 74.73026275634766, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1741839200258255, + "weight_saturation_rate": 0.06122422218322754, + "weight_tensor_scale": 6.480056617874652e-05 + }, + "decoder.transformer_blocks.1.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 131.1938934326172, + "block_scale_min": 32.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.12136677652597427, + "weight_saturation_rate": 0.06213712692260742, + "weight_tensor_scale": 4.515132968663238e-05 + }, + "decoder.transformer_blocks.1.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 63.88050842285156, + "block_scale_min": 16.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.20271283388137817, + "weight_saturation_rate": 0.0610814094543457, + "weight_tensor_scale": 7.541400555055588e-05 + }, + "decoder.transformer_blocks.1.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 100.72238159179688, + "block_scale_min": 26.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.16072425246238708, + "weight_saturation_rate": 0.06214499473571777, + "weight_tensor_scale": 5.979324851068668e-05 + }, + "decoder.transformer_blocks.1.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 59.85573959350586, + "block_scale_min": 11.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.21517379581928253, + "weight_saturation_rate": 0.06285321712493896, + "weight_tensor_scale": 8.004978008102626e-05 + }, + "decoder.transformer_blocks.1.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 88.00189208984375, + "block_scale_min": 12.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1493225246667862, + "weight_saturation_rate": 0.0630413293838501, + "weight_tensor_scale": 5.555153620662168e-05 + }, + "decoder.transformer_blocks.10.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 117.79374694824219, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.181676983833313, + "weight_saturation_rate": 0.0632927417755127, + "weight_tensor_scale": 6.758816743968055e-05 + }, + "decoder.transformer_blocks.10.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 142.12118530273438, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.14077775180339813, + "weight_saturation_rate": 0.06362342834472656, + "weight_tensor_scale": 5.237267760094255e-05 + }, + "decoder.transformer_blocks.10.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 125.5795669555664, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.17257139086723328, + "weight_saturation_rate": 0.0634922981262207, + "weight_tensor_scale": 6.420067074941471e-05 + }, + "decoder.transformer_blocks.10.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 126.27961730957031, + "block_scale_min": 24.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.16732722520828247, + "weight_saturation_rate": 0.06361246109008789, + "weight_tensor_scale": 6.22497100266628e-05 + }, + "decoder.transformer_blocks.10.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 67.20111083984375, + "block_scale_min": 9.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.29687973856925964, + "weight_saturation_rate": 0.06388822197914124, + "weight_tensor_scale": 0.00011044632992707193 + }, + "decoder.transformer_blocks.10.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 125.43810272216797, + "block_scale_min": 15.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1506909281015396, + "weight_saturation_rate": 0.06380116939544678, + "weight_tensor_scale": 5.606061313301325e-05 + }, + "decoder.transformer_blocks.11.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 115.9443130493164, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.18268615007400513, + "weight_saturation_rate": 0.06326699256896973, + "weight_tensor_scale": 6.79635995766148e-05 + }, + "decoder.transformer_blocks.11.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 149.07777404785156, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.13117121160030365, + "weight_saturation_rate": 0.0639338493347168, + "weight_tensor_scale": 4.8798814532347023e-05 + }, + "decoder.transformer_blocks.11.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 109.59060668945312, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.19404767453670502, + "weight_saturation_rate": 0.06316685676574707, + "weight_tensor_scale": 7.219035614980385e-05 + }, + "decoder.transformer_blocks.11.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 153.5157012939453, + "block_scale_min": 32.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.13635319471359253, + "weight_saturation_rate": 0.06336736679077148, + "weight_tensor_scale": 5.072663407190703e-05 + }, + "decoder.transformer_blocks.11.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 77.02044677734375, + "block_scale_min": 14.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.2584809958934784, + "weight_saturation_rate": 0.0638146698474884, + "weight_tensor_scale": 9.616108582122251e-05 + }, + "decoder.transformer_blocks.11.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 132.86219787597656, + "block_scale_min": 14.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.14137504994869232, + "weight_saturation_rate": 0.06363725662231445, + "weight_tensor_scale": 5.2594885346479714e-05 + }, + "decoder.transformer_blocks.12.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 109.84683990478516, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.17352379858493805, + "weight_saturation_rate": 0.06326889991760254, + "weight_tensor_scale": 6.45549880573526e-05 + }, + "decoder.transformer_blocks.12.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 144.46754455566406, + "block_scale_min": 16.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.12776601314544678, + "weight_saturation_rate": 0.0639486312866211, + "weight_tensor_scale": 4.7532001190120354e-05 + }, + "decoder.transformer_blocks.12.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 96.23648071289062, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.2018156796693802, + "weight_saturation_rate": 0.0633537769317627, + "weight_tensor_scale": 7.508024282287806e-05 + }, + "decoder.transformer_blocks.12.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 101.17506408691406, + "block_scale_min": 24.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.19016887247562408, + "weight_saturation_rate": 0.06328678131103516, + "weight_tensor_scale": 7.074735185597092e-05 + }, + "decoder.transformer_blocks.12.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 64.27044677734375, + "block_scale_min": 12.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.309335321187973, + "weight_saturation_rate": 0.06374791264533997, + "weight_tensor_scale": 0.0001150801035691984 + }, + "decoder.transformer_blocks.12.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 113.07392883300781, + "block_scale_min": 12.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.16443794965744019, + "weight_saturation_rate": 0.06346327066421509, + "weight_tensor_scale": 6.117483280831948e-05 + }, + "decoder.transformer_blocks.13.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 114.51170349121094, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.17074373364448547, + "weight_saturation_rate": 0.06324982643127441, + "weight_tensor_scale": 6.352073251036927e-05 + }, + "decoder.transformer_blocks.13.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 124.75965881347656, + "block_scale_min": 16.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.15020859241485596, + "weight_saturation_rate": 0.06390380859375, + "weight_tensor_scale": 5.588117346633226e-05 + }, + "decoder.transformer_blocks.13.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 93.08543395996094, + "block_scale_min": 11.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.21441102027893066, + "weight_saturation_rate": 0.06354093551635742, + "weight_tensor_scale": 7.976600318215787e-05 + }, + "decoder.transformer_blocks.13.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 128.46485900878906, + "block_scale_min": 26.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.15371282398700714, + "weight_saturation_rate": 0.0632939338684082, + "weight_tensor_scale": 5.718483225791715e-05 + }, + "decoder.transformer_blocks.13.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 96.45217895507812, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.20434701442718506, + "weight_saturation_rate": 0.06333529949188232, + "weight_tensor_scale": 7.60219554649666e-05 + }, + "decoder.transformer_blocks.13.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 125.5994873046875, + "block_scale_min": 15.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.14764975011348724, + "weight_saturation_rate": 0.06373924016952515, + "weight_tensor_scale": 5.492922355188057e-05 + }, + "decoder.transformer_blocks.14.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 105.21479034423828, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.18185652792453766, + "weight_saturation_rate": 0.0632016658782959, + "weight_tensor_scale": 6.765496073057875e-05 + }, + "decoder.transformer_blocks.14.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 143.87298583984375, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1290559470653534, + "weight_saturation_rate": 0.0637052059173584, + "weight_tensor_scale": 4.801188697456382e-05 + }, + "decoder.transformer_blocks.14.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 88.98287963867188, + "block_scale_min": 14.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.22003640234470367, + "weight_saturation_rate": 0.06332564353942871, + "weight_tensor_scale": 8.185877959476784e-05 + }, + "decoder.transformer_blocks.14.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 144.05105590820312, + "block_scale_min": 26.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.132658451795578, + "weight_saturation_rate": 0.06318092346191406, + "weight_tensor_scale": 4.935210381518118e-05 + }, + "decoder.transformer_blocks.14.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 106.98713684082031, + "block_scale_min": 15.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1826309710741043, + "weight_saturation_rate": 0.06330016255378723, + "weight_tensor_scale": 6.794307410018519e-05 + }, + "decoder.transformer_blocks.14.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 118.69529724121094, + "block_scale_min": 12.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1567382663488388, + "weight_saturation_rate": 0.06352072954177856, + "weight_tensor_scale": 5.831036833114922e-05 + }, + "decoder.transformer_blocks.15.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 96.12029266357422, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.2069128155708313, + "weight_saturation_rate": 0.06322860717773438, + "weight_tensor_scale": 7.697649562032893e-05 + }, + "decoder.transformer_blocks.15.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 126.98287200927734, + "block_scale_min": 14.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.15120065212249756, + "weight_saturation_rate": 0.0637824535369873, + "weight_tensor_scale": 5.6250242778332904e-05 + }, + "decoder.transformer_blocks.15.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 95.70137786865234, + "block_scale_min": 16.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.2126498967409134, + "weight_saturation_rate": 0.06312322616577148, + "weight_tensor_scale": 7.91108250268735e-05 + }, + "decoder.transformer_blocks.15.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 119.01168823242188, + "block_scale_min": 24.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.16560104489326477, + "weight_saturation_rate": 0.06329464912414551, + "weight_tensor_scale": 6.160753400763497e-05 + }, + "decoder.transformer_blocks.15.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 100.84209442138672, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.19181042909622192, + "weight_saturation_rate": 0.06330886483192444, + "weight_tensor_scale": 7.135804480640218e-05 + }, + "decoder.transformer_blocks.15.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 135.2824249267578, + "block_scale_min": 15.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1372319757938385, + "weight_saturation_rate": 0.06366288661956787, + "weight_tensor_scale": 5.105356467538513e-05 + }, + "decoder.transformer_blocks.16.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 98.16344451904297, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.17273366451263428, + "weight_saturation_rate": 0.06305050849914551, + "weight_tensor_scale": 6.426103936973959e-05 + }, + "decoder.transformer_blocks.16.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 122.8104476928711, + "block_scale_min": 26.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1384819746017456, + "weight_saturation_rate": 0.06263399124145508, + "weight_tensor_scale": 5.151859295438044e-05 + }, + "decoder.transformer_blocks.16.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 100.86898040771484, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1769983321428299, + "weight_saturation_rate": 0.06293416023254395, + "weight_tensor_scale": 6.584759830730036e-05 + }, + "decoder.transformer_blocks.16.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 126.76190185546875, + "block_scale_min": 28.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.13837072253227234, + "weight_saturation_rate": 0.06289100646972656, + "weight_tensor_scale": 5.1477203669492155e-05 + }, + "decoder.transformer_blocks.16.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 98.8709945678711, + "block_scale_min": 16.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1947374790906906, + "weight_saturation_rate": 0.0633736252784729, + "weight_tensor_scale": 7.24469791748561e-05 + }, + "decoder.transformer_blocks.16.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 114.95946502685547, + "block_scale_min": 11.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.16469672322273254, + "weight_saturation_rate": 0.06334656476974487, + "weight_tensor_scale": 6.127110100351274e-05 + }, + "decoder.transformer_blocks.17.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 113.94300842285156, + "block_scale_min": 26.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.13366688787937164, + "weight_saturation_rate": 0.06299161911010742, + "weight_tensor_scale": 4.972726674168371e-05 + }, + "decoder.transformer_blocks.17.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 105.87223052978516, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1499568372964859, + "weight_saturation_rate": 0.06264925003051758, + "weight_tensor_scale": 5.578751370194368e-05 + }, + "decoder.transformer_blocks.17.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 114.89053344726562, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1372968852519989, + "weight_saturation_rate": 0.06267571449279785, + "weight_tensor_scale": 5.10777099407278e-05 + }, + "decoder.transformer_blocks.17.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 89.81871032714844, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.17661544680595398, + "weight_saturation_rate": 0.06339025497436523, + "weight_tensor_scale": 6.570514960912988e-05 + }, + "decoder.transformer_blocks.17.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 94.3211898803711, + "block_scale_min": 15.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.20289990305900574, + "weight_saturation_rate": 0.06336003541946411, + "weight_tensor_scale": 7.548359280917794e-05 + }, + "decoder.transformer_blocks.17.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 108.44267272949219, + "block_scale_min": 13.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.17353297770023346, + "weight_saturation_rate": 0.06325477361679077, + "weight_tensor_scale": 6.455840048147365e-05 + }, + "decoder.transformer_blocks.18.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 94.04303741455078, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.14825467765331268, + "weight_saturation_rate": 0.06333804130554199, + "weight_tensor_scale": 5.515427255886607e-05 + }, + "decoder.transformer_blocks.18.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 138.96444702148438, + "block_scale_min": 48.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.09394112229347229, + "weight_saturation_rate": 0.06767702102661133, + "weight_tensor_scale": 3.494833435979672e-05 + }, + "decoder.transformer_blocks.18.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 86.57354736328125, + "block_scale_min": 16.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.17398805916309357, + "weight_saturation_rate": 0.06301307678222656, + "weight_tensor_scale": 6.472770473919809e-05 + }, + "decoder.transformer_blocks.18.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 131.671875, + "block_scale_min": 36.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.10949879139661789, + "weight_saturation_rate": 0.06285881996154785, + "weight_tensor_scale": 4.073615855304524e-05 + }, + "decoder.transformer_blocks.18.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 76.65528106689453, + "block_scale_min": 12.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.2518253028392792, + "weight_saturation_rate": 0.06360328197479248, + "weight_tensor_scale": 9.368501196149737e-05 + }, + "decoder.transformer_blocks.18.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 130.92648315429688, + "block_scale_min": 14.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1470203548669815, + "weight_saturation_rate": 0.06357890367507935, + "weight_tensor_scale": 5.4695072321919724e-05 + }, + "decoder.transformer_blocks.19.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 133.95481872558594, + "block_scale_min": 36.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.09691120684146881, + "weight_saturation_rate": 0.06311202049255371, + "weight_tensor_scale": 3.605327583500184e-05 + }, + "decoder.transformer_blocks.19.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 93.61833953857422, + "block_scale_min": 30.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.13914045691490173, + "weight_saturation_rate": 0.060611724853515625, + "weight_tensor_scale": 5.176356353331357e-05 + }, + "decoder.transformer_blocks.19.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 114.35924530029297, + "block_scale_min": 26.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.12149816751480103, + "weight_saturation_rate": 0.06265544891357422, + "weight_tensor_scale": 4.520021320786327e-05 + }, + "decoder.transformer_blocks.19.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 127.03355407714844, + "block_scale_min": 36.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.10536051541566849, + "weight_saturation_rate": 0.06217050552368164, + "weight_tensor_scale": 3.9196620491566136e-05 + }, + "decoder.transformer_blocks.19.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 102.34036254882812, + "block_scale_min": 15.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1912178099155426, + "weight_saturation_rate": 0.06347724795341492, + "weight_tensor_scale": 7.11375760147348e-05 + }, + "decoder.transformer_blocks.19.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 141.51687622070312, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.13879787921905518, + "weight_saturation_rate": 0.06348025798797607, + "weight_tensor_scale": 5.1636117859743536e-05 + }, + "decoder.transformer_blocks.2.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 67.20957946777344, + "block_scale_min": 14.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.22118806838989258, + "weight_saturation_rate": 0.0621187686920166, + "weight_tensor_scale": 8.228723163483664e-05 + }, + "decoder.transformer_blocks.2.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 132.078369140625, + "block_scale_min": 30.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.13449425995349884, + "weight_saturation_rate": 0.06348919868469238, + "weight_tensor_scale": 5.003506885259412e-05 + }, + "decoder.transformer_blocks.2.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 75.21023559570312, + "block_scale_min": 15.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.19917799532413483, + "weight_saturation_rate": 0.0621800422668457, + "weight_tensor_scale": 7.409895624732599e-05 + }, + "decoder.transformer_blocks.2.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 136.08908081054688, + "block_scale_min": 26.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.13578961789608002, + "weight_saturation_rate": 0.06307506561279297, + "weight_tensor_scale": 5.051697007729672e-05 + }, + "decoder.transformer_blocks.2.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 63.421104431152344, + "block_scale_min": 13.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.254969984292984, + "weight_saturation_rate": 0.06336939334869385, + "weight_tensor_scale": 9.48549059103243e-05 + }, + "decoder.transformer_blocks.2.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 108.21012878417969, + "block_scale_min": 12.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1491750031709671, + "weight_saturation_rate": 0.0635833740234375, + "weight_tensor_scale": 5.549665365833789e-05 + }, + "decoder.transformer_blocks.20.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 140.6821746826172, + "block_scale_min": 36.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.08891433477401733, + "weight_saturation_rate": 0.06363630294799805, + "weight_tensor_scale": 3.307824954390526e-05 + }, + "decoder.transformer_blocks.20.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 116.19229125976562, + "block_scale_min": 32.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.11038126051425934, + "weight_saturation_rate": 0.0617680549621582, + "weight_tensor_scale": 4.106445703655481e-05 + }, + "decoder.transformer_blocks.20.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 95.84339141845703, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.13798609375953674, + "weight_saturation_rate": 0.06282281875610352, + "weight_tensor_scale": 5.1334114687051624e-05 + }, + "decoder.transformer_blocks.20.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 137.06500244140625, + "block_scale_min": 36.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.09575022011995316, + "weight_saturation_rate": 0.06285405158996582, + "weight_tensor_scale": 3.562136407708749e-05 + }, + "decoder.transformer_blocks.20.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 88.50335693359375, + "block_scale_min": 12.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.21761423349380493, + "weight_saturation_rate": 0.06361514329910278, + "weight_tensor_scale": 8.095767407212406e-05 + }, + "decoder.transformer_blocks.20.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 137.79965209960938, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1416510045528412, + "weight_saturation_rate": 0.06350356340408325, + "weight_tensor_scale": 5.269754910841584e-05 + }, + "decoder.transformer_blocks.21.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 122.42832946777344, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.10724260658025742, + "weight_saturation_rate": 0.06305789947509766, + "weight_tensor_scale": 3.989680408267304e-05 + }, + "decoder.transformer_blocks.21.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 101.33067321777344, + "block_scale_min": 32.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.13749203085899353, + "weight_saturation_rate": 0.06086468696594238, + "weight_tensor_scale": 5.115030944580212e-05 + }, + "decoder.transformer_blocks.21.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 96.2477798461914, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1423412561416626, + "weight_saturation_rate": 0.06315493583679199, + "weight_tensor_scale": 5.295433584251441e-05 + }, + "decoder.transformer_blocks.21.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 99.03059387207031, + "block_scale_min": 24.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.14650732278823853, + "weight_saturation_rate": 0.06248307228088379, + "weight_tensor_scale": 5.450421303976327e-05 + }, + "decoder.transformer_blocks.21.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 86.46150970458984, + "block_scale_min": 13.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.2180563062429428, + "weight_saturation_rate": 0.06357273459434509, + "weight_tensor_scale": 8.112213981803507e-05 + }, + "decoder.transformer_blocks.21.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 90.38046264648438, + "block_scale_min": 11.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.21643376350402832, + "weight_saturation_rate": 0.06327265501022339, + "weight_tensor_scale": 8.051851182244718e-05 + }, + "decoder.transformer_blocks.22.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 122.40064239501953, + "block_scale_min": 30.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.10354143381118774, + "weight_saturation_rate": 0.06300878524780273, + "weight_tensor_scale": 3.851988003589213e-05 + }, + "decoder.transformer_blocks.22.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 122.74874877929688, + "block_scale_min": 36.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.10274393856525421, + "weight_saturation_rate": 0.0644228458404541, + "weight_tensor_scale": 3.822319195023738e-05 + }, + "decoder.transformer_blocks.22.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 80.0743408203125, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.16341716051101685, + "weight_saturation_rate": 0.06317520141601562, + "weight_tensor_scale": 6.0795075114583597e-05 + }, + "decoder.transformer_blocks.22.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 122.39105987548828, + "block_scale_min": 30.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.10349239408969879, + "weight_saturation_rate": 0.06182432174682617, + "weight_tensor_scale": 3.8501635572174564e-05 + }, + "decoder.transformer_blocks.22.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 92.5489273071289, + "block_scale_min": 14.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.19356675446033478, + "weight_saturation_rate": 0.06368228793144226, + "weight_tensor_scale": 7.201144035207108e-05 + }, + "decoder.transformer_blocks.22.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 93.48477172851562, + "block_scale_min": 11.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.2033165991306305, + "weight_saturation_rate": 0.06312745809555054, + "weight_tensor_scale": 7.563861436210573e-05 + }, + "decoder.transformer_blocks.23.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 63.12794876098633, + "block_scale_min": 13.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.2226472795009613, + "weight_saturation_rate": 0.06331753730773926, + "weight_tensor_scale": 8.283009083243087e-05 + }, + "decoder.transformer_blocks.23.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 84.12667846679688, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.17239458858966827, + "weight_saturation_rate": 0.06477689743041992, + "weight_tensor_scale": 6.41348960925825e-05 + }, + "decoder.transformer_blocks.23.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 88.04548645019531, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.16747435927391052, + "weight_saturation_rate": 0.06301379203796387, + "weight_tensor_scale": 6.23044470557943e-05 + }, + "decoder.transformer_blocks.23.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 106.48403930664062, + "block_scale_min": 24.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.14637602865695953, + "weight_saturation_rate": 0.06256961822509766, + "weight_tensor_scale": 5.445536953629926e-05 + }, + "decoder.transformer_blocks.23.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 76.31051635742188, + "block_scale_min": 11.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.21919091045856476, + "weight_saturation_rate": 0.06347888708114624, + "weight_tensor_scale": 8.154423994710669e-05 + }, + "decoder.transformer_blocks.23.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 131.80198669433594, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1354273408651352, + "weight_saturation_rate": 0.06329488754272461, + "weight_tensor_scale": 5.03821975144092e-05 + }, + "decoder.transformer_blocks.24.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 135.21517944335938, + "block_scale_min": 32.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.09074429422616959, + "weight_saturation_rate": 0.06375527381896973, + "weight_tensor_scale": 3.375903906999156e-05 + }, + "decoder.transformer_blocks.24.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 103.29702758789062, + "block_scale_min": 32.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.12124158442020416, + "weight_saturation_rate": 0.05776834487915039, + "weight_tensor_scale": 4.510475628194399e-05 + }, + "decoder.transformer_blocks.24.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 75.72248840332031, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.15767353773117065, + "weight_saturation_rate": 0.06397438049316406, + "weight_tensor_scale": 5.8658311900217086e-05 + }, + "decoder.transformer_blocks.24.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 134.26473999023438, + "block_scale_min": 36.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.0842275321483612, + "weight_saturation_rate": 0.06291913986206055, + "weight_tensor_scale": 3.1334646337199956e-05 + }, + "decoder.transformer_blocks.24.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 71.41278076171875, + "block_scale_min": 10.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.20589236915111542, + "weight_saturation_rate": 0.06330448389053345, + "weight_tensor_scale": 7.65968652558513e-05 + }, + "decoder.transformer_blocks.24.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 98.04750061035156, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.15733633935451508, + "weight_saturation_rate": 0.0624733567237854, + "weight_tensor_scale": 5.853286711499095e-05 + }, + "decoder.transformer_blocks.25.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 105.37522888183594, + "block_scale_min": 24.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.12656976282596588, + "weight_saturation_rate": 0.0629279613494873, + "weight_tensor_scale": 4.7086967242648825e-05 + }, + "decoder.transformer_blocks.25.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 91.49143981933594, + "block_scale_min": 32.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.14189814031124115, + "weight_saturation_rate": 0.062224388122558594, + "weight_tensor_scale": 5.2789488108828664e-05 + }, + "decoder.transformer_blocks.25.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 101.6304931640625, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1390058696269989, + "weight_saturation_rate": 0.06282854080200195, + "weight_tensor_scale": 5.171349403099157e-05 + }, + "decoder.transformer_blocks.25.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 114.55616760253906, + "block_scale_min": 30.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.11830809712409973, + "weight_saturation_rate": 0.06175971031188965, + "weight_tensor_scale": 4.4013428123435006e-05 + }, + "decoder.transformer_blocks.25.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 88.21012878417969, + "block_scale_min": 16.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.14805448055267334, + "weight_saturation_rate": 0.0629357397556305, + "weight_tensor_scale": 5.507979221874848e-05 + }, + "decoder.transformer_blocks.25.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 70.44404602050781, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.150229811668396, + "weight_saturation_rate": 0.0611761212348938, + "weight_tensor_scale": 5.5889067880343646e-05 + }, + "decoder.transformer_blocks.26.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 105.23062133789062, + "block_scale_min": 24.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1165730357170105, + "weight_saturation_rate": 0.06331419944763184, + "weight_tensor_scale": 4.336794518167153e-05 + }, + "decoder.transformer_blocks.26.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 129.78099060058594, + "block_scale_min": 44.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.09884219616651535, + "weight_saturation_rate": 0.06684184074401855, + "weight_tensor_scale": 3.6771649320144206e-05 + }, + "decoder.transformer_blocks.26.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 72.18895721435547, + "block_scale_min": 16.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.18377356231212616, + "weight_saturation_rate": 0.06288385391235352, + "weight_tensor_scale": 6.83681428199634e-05 + }, + "decoder.transformer_blocks.26.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 126.05873107910156, + "block_scale_min": 36.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1090700626373291, + "weight_saturation_rate": 0.06193733215332031, + "weight_tensor_scale": 4.057666228618473e-05 + }, + "decoder.transformer_blocks.26.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 79.20211791992188, + "block_scale_min": 12.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.21352577209472656, + "weight_saturation_rate": 0.06346288323402405, + "weight_tensor_scale": 7.943667151266709e-05 + }, + "decoder.transformer_blocks.26.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 88.44673156738281, + "block_scale_min": 14.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.20182697474956512, + "weight_saturation_rate": 0.06303489208221436, + "weight_tensor_scale": 7.508444105042145e-05 + }, + "decoder.transformer_blocks.27.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 144.62693786621094, + "block_scale_min": 44.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.0808972492814064, + "weight_saturation_rate": 0.06395673751831055, + "weight_tensor_scale": 3.0095703550614417e-05 + }, + "decoder.transformer_blocks.27.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 184.20126342773438, + "block_scale_min": 52.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.06756629049777985, + "weight_saturation_rate": 0.0604853630065918, + "weight_tensor_scale": 2.5136268959613517e-05 + }, + "decoder.transformer_blocks.27.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 82.7099380493164, + "block_scale_min": 24.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.14070560038089752, + "weight_saturation_rate": 0.06344795227050781, + "weight_tensor_scale": 5.234583295532502e-05 + }, + "decoder.transformer_blocks.27.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 124.93556213378906, + "block_scale_min": 40.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.08558806031942368, + "weight_saturation_rate": 0.06191849708557129, + "weight_tensor_scale": 3.1840798328630626e-05 + }, + "decoder.transformer_blocks.27.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 81.75971984863281, + "block_scale_min": 13.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.2053391933441162, + "weight_saturation_rate": 0.06343582272529602, + "weight_tensor_scale": 7.639107207069173e-05 + }, + "decoder.transformer_blocks.27.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 83.60118103027344, + "block_scale_min": 13.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.2100093811750412, + "weight_saturation_rate": 0.06316208839416504, + "weight_tensor_scale": 7.812849071342498e-05 + }, + "decoder.transformer_blocks.28.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 62.01121520996094, + "block_scale_min": 14.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.17582416534423828, + "weight_saturation_rate": 0.06314396858215332, + "weight_tensor_scale": 6.541077891597524e-05 + }, + "decoder.transformer_blocks.28.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 195.05764770507812, + "block_scale_min": 48.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.06364677846431732, + "weight_saturation_rate": 0.0613560676574707, + "weight_tensor_scale": 2.3678117941017263e-05 + }, + "decoder.transformer_blocks.28.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 60.885986328125, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.17778849601745605, + "weight_saturation_rate": 0.06265521049499512, + "weight_tensor_scale": 6.614155427087098e-05 + }, + "decoder.transformer_blocks.28.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 170.52200317382812, + "block_scale_min": 48.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.06753546744585037, + "weight_saturation_rate": 0.06304502487182617, + "weight_tensor_scale": 2.5124802050413564e-05 + }, + "decoder.transformer_blocks.28.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 85.31277465820312, + "block_scale_min": 13.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.17417527735233307, + "weight_saturation_rate": 0.06341993808746338, + "weight_tensor_scale": 6.479735020548105e-05 + }, + "decoder.transformer_blocks.28.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 87.6125717163086, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1788119375705719, + "weight_saturation_rate": 0.06271052360534668, + "weight_tensor_scale": 6.652229785686359e-05 + }, + "decoder.transformer_blocks.29.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 120.63053894042969, + "block_scale_min": 28.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.10001787543296814, + "weight_saturation_rate": 0.06273150444030762, + "weight_tensor_scale": 3.720903259818442e-05 + }, + "decoder.transformer_blocks.29.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 105.4293212890625, + "block_scale_min": 36.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.12382815778255463, + "weight_saturation_rate": 0.06033587455749512, + "weight_tensor_scale": 4.606702350429259e-05 + }, + "decoder.transformer_blocks.29.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 108.61454010009766, + "block_scale_min": 28.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.11522454768419266, + "weight_saturation_rate": 0.0625905990600586, + "weight_tensor_scale": 4.28662751801312e-05 + }, + "decoder.transformer_blocks.29.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 122.14743041992188, + "block_scale_min": 28.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.10934911668300629, + "weight_saturation_rate": 0.061919212341308594, + "weight_tensor_scale": 4.06804756494239e-05 + }, + "decoder.transformer_blocks.29.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 80.64563751220703, + "block_scale_min": 15.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.14406971633434296, + "weight_saturation_rate": 0.06302440166473389, + "weight_tensor_scale": 5.3597366786561906e-05 + }, + "decoder.transformer_blocks.29.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 57.13140106201172, + "block_scale_min": 14.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.19090536236763, + "weight_saturation_rate": 0.06206321716308594, + "weight_tensor_scale": 7.102134259184822e-05 + }, + "decoder.transformer_blocks.3.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 93.24532318115234, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.18933217227458954, + "weight_saturation_rate": 0.06301546096801758, + "weight_tensor_scale": 7.043607911327854e-05 + }, + "decoder.transformer_blocks.3.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 173.57687377929688, + "block_scale_min": 30.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.11281410604715347, + "weight_saturation_rate": 0.06317830085754395, + "weight_tensor_scale": 4.196953523205593e-05 + }, + "decoder.transformer_blocks.3.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 82.83694458007812, + "block_scale_min": 15.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.21683627367019653, + "weight_saturation_rate": 0.06337594985961914, + "weight_tensor_scale": 8.066825830610469e-05 + }, + "decoder.transformer_blocks.3.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 147.6359100341797, + "block_scale_min": 28.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1352991908788681, + "weight_saturation_rate": 0.06335783004760742, + "weight_tensor_scale": 5.033452180214226e-05 + }, + "decoder.transformer_blocks.3.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 88.52576446533203, + "block_scale_min": 16.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.2027779519557953, + "weight_saturation_rate": 0.06342566013336182, + "weight_tensor_scale": 7.54382272134535e-05 + }, + "decoder.transformer_blocks.3.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 80.6900405883789, + "block_scale_min": 10.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.2251390814781189, + "weight_saturation_rate": 0.06315171718597412, + "weight_tensor_scale": 8.375709876418114e-05 + }, + "decoder.transformer_blocks.30.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 112.25161743164062, + "block_scale_min": 28.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1009037122130394, + "weight_saturation_rate": 0.06269192695617676, + "weight_tensor_scale": 3.753858254640363e-05 + }, + "decoder.transformer_blocks.30.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 155.5049591064453, + "block_scale_min": 52.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.08013541996479034, + "weight_saturation_rate": 0.06705045700073242, + "weight_tensor_scale": 2.9812284992658533e-05 + }, + "decoder.transformer_blocks.30.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 110.75624084472656, + "block_scale_min": 24.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.11072496324777603, + "weight_saturation_rate": 0.06255912780761719, + "weight_tensor_scale": 4.119232471566647e-05 + }, + "decoder.transformer_blocks.30.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 141.43759155273438, + "block_scale_min": 44.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.08750093728303909, + "weight_saturation_rate": 0.06305909156799316, + "weight_tensor_scale": 3.255243427702226e-05 + }, + "decoder.transformer_blocks.30.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 59.12864685058594, + "block_scale_min": 13.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1875629872083664, + "weight_saturation_rate": 0.06272551417350769, + "weight_tensor_scale": 6.97778959874995e-05 + }, + "decoder.transformer_blocks.30.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 63.41782760620117, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.14543378353118896, + "weight_saturation_rate": 0.06066638231277466, + "weight_tensor_scale": 5.4104832088341936e-05 + }, + "decoder.transformer_blocks.31.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 113.39540100097656, + "block_scale_min": 32.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.10184811800718307, + "weight_saturation_rate": 0.06215047836303711, + "weight_tensor_scale": 3.788992398767732e-05 + }, + "decoder.transformer_blocks.31.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 275.735595703125, + "block_scale_min": 72.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.044625747948884964, + "weight_saturation_rate": 0.07172203063964844, + "weight_tensor_scale": 1.6601841707597487e-05 + }, + "decoder.transformer_blocks.31.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 69.46107482910156, + "block_scale_min": 16.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1779913455247879, + "weight_saturation_rate": 0.0626683235168457, + "weight_tensor_scale": 6.62170205032453e-05 + }, + "decoder.transformer_blocks.31.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 107.52058410644531, + "block_scale_min": 32.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.10288650542497635, + "weight_saturation_rate": 0.06027698516845703, + "weight_tensor_scale": 3.827623004326597e-05 + }, + "decoder.transformer_blocks.31.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 65.64836120605469, + "block_scale_min": 13.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.15962685644626617, + "weight_saturation_rate": 0.06254985928535461, + "weight_tensor_scale": 5.938499089097604e-05 + }, + "decoder.transformer_blocks.31.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 69.52937316894531, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.13101471960544586, + "weight_saturation_rate": 0.06510734558105469, + "weight_tensor_scale": 4.8740595957497135e-05 + }, + "decoder.transformer_blocks.32.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 110.68551635742188, + "block_scale_min": 28.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1004888117313385, + "weight_saturation_rate": 0.06215167045593262, + "weight_tensor_scale": 3.7384230381576344e-05 + }, + "decoder.transformer_blocks.32.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 305.65972900390625, + "block_scale_min": 96.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.04033840075135231, + "weight_saturation_rate": 0.07170343399047852, + "weight_tensor_scale": 1.5006845387688372e-05 + }, + "decoder.transformer_blocks.32.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 62.7059440612793, + "block_scale_min": 15.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1865345686674118, + "weight_saturation_rate": 0.062042951583862305, + "weight_tensor_scale": 6.93953043082729e-05 + }, + "decoder.transformer_blocks.32.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 109.6470718383789, + "block_scale_min": 30.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.10024440288543701, + "weight_saturation_rate": 0.05968809127807617, + "weight_tensor_scale": 3.72933063772507e-05 + }, + "decoder.transformer_blocks.32.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 78.35759735107422, + "block_scale_min": 16.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1355494260787964, + "weight_saturation_rate": 0.06294971704483032, + "weight_tensor_scale": 5.042761404183693e-05 + }, + "decoder.transformer_blocks.32.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 67.98856353759766, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.13304503262043, + "weight_saturation_rate": 0.06481021642684937, + "weight_tensor_scale": 4.949592039338313e-05 + }, + "decoder.transformer_blocks.33.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 85.7637939453125, + "block_scale_min": 24.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.12888824939727783, + "weight_saturation_rate": 0.06394076347351074, + "weight_tensor_scale": 4.7949499276001006e-05 + }, + "decoder.transformer_blocks.33.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 123.85324096679688, + "block_scale_min": 44.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.09988239407539368, + "weight_saturation_rate": 0.07051515579223633, + "weight_tensor_scale": 3.7158628401812166e-05 + }, + "decoder.transformer_blocks.33.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 68.52757263183594, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.17321066558361053, + "weight_saturation_rate": 0.06219673156738281, + "weight_tensor_scale": 6.443849269999191e-05 + }, + "decoder.transformer_blocks.33.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 68.66661071777344, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.17006415128707886, + "weight_saturation_rate": 0.06168818473815918, + "weight_tensor_scale": 6.326791481114924e-05 + }, + "decoder.transformer_blocks.33.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 67.36929321289062, + "block_scale_min": 12.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.153999924659729, + "weight_saturation_rate": 0.062358319759368896, + "weight_tensor_scale": 5.7291639677714556e-05 + }, + "decoder.transformer_blocks.33.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 44.803932189941406, + "block_scale_min": 13.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1998044103384018, + "weight_saturation_rate": 0.06494921445846558, + "weight_tensor_scale": 7.433199789375067e-05 + }, + "decoder.transformer_blocks.34.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 100.22808837890625, + "block_scale_min": 30.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.10355579107999802, + "weight_saturation_rate": 0.06267023086547852, + "weight_tensor_scale": 3.852522058878094e-05 + }, + "decoder.transformer_blocks.34.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 220.49786376953125, + "block_scale_min": 48.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.05685105174779892, + "weight_saturation_rate": 0.06061363220214844, + "weight_tensor_scale": 2.1149944586795755e-05 + }, + "decoder.transformer_blocks.34.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 100.21388244628906, + "block_scale_min": 26.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1136665791273117, + "weight_saturation_rate": 0.06305885314941406, + "weight_tensor_scale": 4.228667603456415e-05 + }, + "decoder.transformer_blocks.34.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 116.71891784667969, + "block_scale_min": 32.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.10370064526796341, + "weight_saturation_rate": 0.06025886535644531, + "weight_tensor_scale": 3.857910996885039e-05 + }, + "decoder.transformer_blocks.34.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 54.74583053588867, + "block_scale_min": 12.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.17882749438285828, + "weight_saturation_rate": 0.06258809566497803, + "weight_tensor_scale": 6.652808951912448e-05 + }, + "decoder.transformer_blocks.34.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 63.4363899230957, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1357102394104004, + "weight_saturation_rate": 0.06485897302627563, + "weight_tensor_scale": 5.0487440603319556e-05 + }, + "decoder.transformer_blocks.35.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 98.82254791259766, + "block_scale_min": 26.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.10737597197294235, + "weight_saturation_rate": 0.06316637992858887, + "weight_tensor_scale": 3.994641883764416e-05 + }, + "decoder.transformer_blocks.35.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 139.25054931640625, + "block_scale_min": 52.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.08949720114469528, + "weight_saturation_rate": 0.07316160202026367, + "weight_tensor_scale": 3.329509127070196e-05 + }, + "decoder.transformer_blocks.35.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 87.60562133789062, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.13313177227973938, + "weight_saturation_rate": 0.06335616111755371, + "weight_tensor_scale": 4.9528189265402034e-05 + }, + "decoder.transformer_blocks.35.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 49.58196258544922, + "block_scale_min": 12.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.260823130607605, + "weight_saturation_rate": 0.06226515769958496, + "weight_tensor_scale": 9.703241812530905e-05 + }, + "decoder.transformer_blocks.35.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 22.78741455078125, + "block_scale_min": 4.5, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.44047603011131287, + "weight_saturation_rate": 0.06317579746246338, + "weight_tensor_scale": 0.0001638675748836249 + }, + "decoder.transformer_blocks.35.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 39.70322036743164, + "block_scale_min": 10.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.2142769992351532, + "weight_saturation_rate": 0.06896787881851196, + "weight_tensor_scale": 7.971614832058549e-05 + }, + "decoder.transformer_blocks.4.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 79.93778228759766, + "block_scale_min": 13.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.2578102946281433, + "weight_saturation_rate": 0.06357169151306152, + "weight_tensor_scale": 9.591157140675932e-05 + }, + "decoder.transformer_blocks.4.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 168.37149047851562, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.12235993146896362, + "weight_saturation_rate": 0.06327676773071289, + "weight_tensor_scale": 4.552081009023823e-05 + }, + "decoder.transformer_blocks.4.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 100.82862091064453, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.2072661966085434, + "weight_saturation_rate": 0.06325030326843262, + "weight_tensor_scale": 7.710796489845961e-05 + }, + "decoder.transformer_blocks.4.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 143.56973266601562, + "block_scale_min": 30.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1447458416223526, + "weight_saturation_rate": 0.0635535717010498, + "weight_tensor_scale": 5.3848900279263034e-05 + }, + "decoder.transformer_blocks.4.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 85.2594223022461, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.22712570428848267, + "weight_saturation_rate": 0.06354731321334839, + "weight_tensor_scale": 8.449616871075705e-05 + }, + "decoder.transformer_blocks.4.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 139.528564453125, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.13606232404708862, + "weight_saturation_rate": 0.06383419036865234, + "weight_tensor_scale": 5.061842603026889e-05 + }, + "decoder.transformer_blocks.5.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 117.29517364501953, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.18505895137786865, + "weight_saturation_rate": 0.06367826461791992, + "weight_tensor_scale": 6.884634058224037e-05 + }, + "decoder.transformer_blocks.5.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 172.8443603515625, + "block_scale_min": 24.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.12227087467908859, + "weight_saturation_rate": 0.06334495544433594, + "weight_tensor_scale": 4.5487679017242044e-05 + }, + "decoder.transformer_blocks.5.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 107.47306060791016, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.20340308547019958, + "weight_saturation_rate": 0.06351566314697266, + "weight_tensor_scale": 7.567078864667565e-05 + }, + "decoder.transformer_blocks.5.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 164.88661193847656, + "block_scale_min": 40.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.13034163415431976, + "weight_saturation_rate": 0.0631861686706543, + "weight_tensor_scale": 4.849019387620501e-05 + }, + "decoder.transformer_blocks.5.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 85.5090103149414, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.22868850827217102, + "weight_saturation_rate": 0.0635601282119751, + "weight_tensor_scale": 8.50775686558336e-05 + }, + "decoder.transformer_blocks.5.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 126.8505630493164, + "block_scale_min": 12.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.15310969948768616, + "weight_saturation_rate": 0.06370985507965088, + "weight_tensor_scale": 5.696045627701096e-05 + }, + "decoder.transformer_blocks.6.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 125.30074310302734, + "block_scale_min": 24.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.17383921146392822, + "weight_saturation_rate": 0.06356358528137207, + "weight_tensor_scale": 6.467232742579654e-05 + }, + "decoder.transformer_blocks.6.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 160.10870361328125, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1295359879732132, + "weight_saturation_rate": 0.06343984603881836, + "weight_tensor_scale": 4.819047171622515e-05 + }, + "decoder.transformer_blocks.6.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 132.66688537597656, + "block_scale_min": 26.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.16644994914531708, + "weight_saturation_rate": 0.06371593475341797, + "weight_tensor_scale": 6.19233469478786e-05 + }, + "decoder.transformer_blocks.6.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 151.74124145507812, + "block_scale_min": 36.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.14564310014247894, + "weight_saturation_rate": 0.0634312629699707, + "weight_tensor_scale": 5.4182703024707735e-05 + }, + "decoder.transformer_blocks.6.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 79.73844909667969, + "block_scale_min": 15.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.25261378288269043, + "weight_saturation_rate": 0.06375506520271301, + "weight_tensor_scale": 9.397834219271317e-05 + }, + "decoder.transformer_blocks.6.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 127.80899047851562, + "block_scale_min": 13.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.15278583765029907, + "weight_saturation_rate": 0.06384968757629395, + "weight_tensor_scale": 5.683997005689889e-05 + }, + "decoder.transformer_blocks.7.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 106.27161407470703, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.20927439630031586, + "weight_saturation_rate": 0.06336569786071777, + "weight_tensor_scale": 7.785506022628397e-05 + }, + "decoder.transformer_blocks.7.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 171.13031005859375, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.12386759370565414, + "weight_saturation_rate": 0.06345558166503906, + "weight_tensor_scale": 4.608169547282159e-05 + }, + "decoder.transformer_blocks.7.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 111.00055694580078, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.20299175381660461, + "weight_saturation_rate": 0.06352353096008301, + "weight_tensor_scale": 7.551776798209175e-05 + }, + "decoder.transformer_blocks.7.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 139.30337524414062, + "block_scale_min": 28.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.16128982603549957, + "weight_saturation_rate": 0.06380033493041992, + "weight_tensor_scale": 6.0003658290952444e-05 + }, + "decoder.transformer_blocks.7.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 96.43820190429688, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.20917943120002747, + "weight_saturation_rate": 0.06332641839981079, + "weight_tensor_scale": 7.781972817610949e-05 + }, + "decoder.transformer_blocks.7.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 137.05520629882812, + "block_scale_min": 16.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.14209486544132233, + "weight_saturation_rate": 0.06361287832260132, + "weight_tensor_scale": 5.286267332849093e-05 + }, + "decoder.transformer_blocks.8.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 85.3823471069336, + "block_scale_min": 18.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.26177093386650085, + "weight_saturation_rate": 0.06327104568481445, + "weight_tensor_scale": 9.738501830724999e-05 + }, + "decoder.transformer_blocks.8.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 166.85250854492188, + "block_scale_min": 22.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.12719978392124176, + "weight_saturation_rate": 0.06348371505737305, + "weight_tensor_scale": 4.7321347665274516e-05 + }, + "decoder.transformer_blocks.8.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 118.80672454833984, + "block_scale_min": 24.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.18902170658111572, + "weight_saturation_rate": 0.0635077953338623, + "weight_tensor_scale": 7.032057328615338e-05 + }, + "decoder.transformer_blocks.8.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 154.32521057128906, + "block_scale_min": 36.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.1448746621608734, + "weight_saturation_rate": 0.06362128257751465, + "weight_tensor_scale": 5.3896823374088854e-05 + }, + "decoder.transformer_blocks.8.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 73.51123046875, + "block_scale_min": 12.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.2722529172897339, + "weight_saturation_rate": 0.06387752294540405, + "weight_tensor_scale": 0.0001012845677905716 + }, + "decoder.transformer_blocks.8.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 109.78668212890625, + "block_scale_min": 11.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.176651731133461, + "weight_saturation_rate": 0.06354117393493652, + "weight_tensor_scale": 6.57186537864618e-05 + }, + "decoder.transformer_blocks.9.attn.to_k.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 100.08259582519531, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.22780494391918182, + "weight_saturation_rate": 0.06336355209350586, + "weight_tensor_scale": 8.474886271869764e-05 + }, + "decoder.transformer_blocks.9.attn.to_out.0.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 157.62557983398438, + "block_scale_min": 20.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.13329845666885376, + "weight_saturation_rate": 0.06349992752075195, + "weight_tensor_scale": 4.959020225214772e-05 + }, + "decoder.transformer_blocks.9.attn.to_q.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 138.57882690429688, + "block_scale_min": 30.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.16424192488193512, + "weight_saturation_rate": 0.0637516975402832, + "weight_tensor_scale": 6.110190588515252e-05 + }, + "decoder.transformer_blocks.9.attn.to_v.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 151.74652099609375, + "block_scale_min": 30.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.14871032536029816, + "weight_saturation_rate": 0.06370306015014648, + "weight_tensor_scale": 5.532378418138251e-05 + }, + "decoder.transformer_blocks.9.ff.net.0.proj.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 61.11538314819336, + "block_scale_min": 10.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 16384, + 2048 + ], + "source_dtype": "float32", + "weight_global_amax": 0.32586395740509033, + "weight_saturation_rate": 0.06368157267570496, + "weight_tensor_scale": 0.00012122915359213948 + }, + "decoder.transformer_blocks.9.ff.net.2.weight": { + "block_scale_max": 448.0, + "block_scale_mean": 109.49793243408203, + "block_scale_min": 11.0, + "calibration": "static_mse_fp8_e4m3_scale_sweep_126_candidates", + "shape": [ + 2048, + 8192 + ], + "source_dtype": "float32", + "weight_global_amax": 0.17389443516731262, + "weight_saturation_rate": 0.06341946125030518, + "weight_tensor_scale": 6.469287473009899e-05 + } + } +}