{ "attention": { "backend": "VSA_QAT_TRAIN_C128", "compression_gates": "trained", "cube_shape": [ 4, 4, 8 ], "kind": "vsa", "physical_block_tokens": 64, "precision": { "coarse": "BF16", "compute": "BF16", "fine": "NVFP4 fake-quant with BF16 matmul", "master_weights": "FP32", "pv_fp8": false, "quant_p": true, "quant_v": true, "quantization_aware": true, "quantize_coarse": false, "quantize_fine": true, "two_level_p": false, "use_global_sf_p": false, "use_global_sf_qkv": false }, "selection": { "effective_topk": "max(1, min(ceil((1-sparsity)*num_blocks), num_blocks))", "mode": "top-p", "sparsity": 0.9 }, "training_kernel_family": "nanovideo_triton_sparse_bf16_mm" }, "checkpoint": { "created_at": "2026-09-23T00:11:19.743546+00:00", "id": "ckpt_26b5969dab50472180706ef42d2aef08", "state": "training", "step": 1000 }, "checkpoint_state": "inference_export", "code_provenance": { "commit": "fefc437be38d0e1c3e231be5745b2d99af5446f6", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/home/hal-ywu/git/vsqa/outputs/training-source-fefc437-mixed600-c128-no-manifest-hash" } }, "config": { "callbacks": { "grad_clip": { "max_grad_norm": 1.0 }, "validation": { "_target_": "fastvideo.train.methods.ode_init.dmd_validation.ShiftedDMDValidationCallback", "attn_qat_infer": false, "dataset_file": "/home/hal-ywu/git/vsqa/outputs/ode-v2-k3-260920/validation_16.json", "every_steps": 50, "guidance_scale": 1.0, "height": 448, "num_frames": 61, "offload_training_state": true, "pipeline_target": "fastvideo.pipelines.basic.wan.wan_shifted_dmd_pipeline.WanShiftedDMDPipeline", "run_at_start": false, "sampling_steps": [ 3 ], "unload_pipeline_after_validation": true, "use_validation_media_conditioning": false, "width": 896 } }, "method": { "_target_": "fastvideo.train.methods.distribution_matching.dmd2.DMD2Method", "dmd_denoising_steps": [ 1000.0, 941.1763916015625, 800.0 ], "fake_score_betas": [ 0.0, 0.999 ], "fake_score_learning_rate": 4e-07, "fake_score_lr_scheduler": "constant", "generator_update_interval": 5, "max_timestep_ratio": 0.98, "min_timestep_ratio": 0.02, "real_score_guidance_scale": 3.5, "rollout_mode": "simulate", "shifted_ode_schedule": { "flow_shift": 8.0, "indices": [ 0, 16, 32 ], "sigmas": [ 1.0, 0.9411764144897461, 0.800000011920929 ], "student_steps": 3, "teacher_steps": 48, "timesteps": [ 1000.0, 941.1763916015625, 800.0 ], "version": 1 }, "warp_denoising_step": false }, "models": { "critic": { "_target_": "fastvideo.train.models.wan.WanModel", "attention_backend": "TORCH_SDPA", "disable_custom_init_weights": true, "enable_gradient_checkpointing_type": "full", "flow_shift": 8.0, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "trainable": true }, "student": { "_target_": "fastvideo.train.models.wan.WanModel", "attention_backend": "VSA_QAT_TRAIN_C128", "enable_gradient_checkpointing_type": "full", "flow_shift": 8.0, "init_from": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0352-weighted-reconode-c128-k3-b16-1000-260922-104437/export-step500", "trainable": true }, "teacher": { "_target_": "fastvideo.train.models.wan.WanModel", "attention_backend": "TORCH_SDPA", "disable_custom_init_weights": true, "flow_shift": 8.0, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "trainable": false } }, "pipeline": { "dit_config": { "quant_config": "nvfp4_qat_train" }, "dmd_denoising_steps": [ 1000.0, 941.1763916015625, 800.0 ], "flow_shift": 8.0, "warp_denoising_step": false }, "training": { "checkpoint": { "checkpoints_total_limit": 0, "output_dir": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0354-dmd-c128-k3-b16-1000-260922-104437/checkpoints", "resume_from_checkpoint": "", "training_state_checkpointing_steps": 50 }, "data": { "data_path": "/home/hal-ywu/git/vsqa/data/mixed/training_dataset", "dataloader_num_workers": 0, "num_frames": 61, "num_height": 448, "num_latent_t": 16, "num_width": 896, "preprocessed_data_type": "text_only", "seed": 1000, "train_batch_size": 4, "training_cfg_rate": 0.0 }, "distributed": { "hsdp_replicate_dim": 4, "hsdp_shard_dim": 1, "num_gpus": 4, "sp_size": 1, "tp_size": 1 }, "dit_precision": "fp32", "loop": { "gradient_accumulation_steps": 1, "max_train_steps": 1000 }, "optimizer": { "betas": [ 0.0, 0.999 ], "learning_rate": 2e-06, "lr_scheduler": "constant", "lr_warmup_steps": 0, "weight_decay": 0.01 }, "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0354-dmd-c128-k3-b16-1000-260922-104437", "trackers": [ "wandb" ] }, "vsa": { "cache_tile_buf": false, "sparsity": 0.9 } } }, "config_sha256": { "resolved": "7047bb89d1c6f40920c0f438f59ff775efe7a8c175212d0ff79c92a379d5af2d", "source": "dfeda6ad9b95acea42d430f87b7017a8daf60489471284ddd9ede39aaf4b14e7" }, "metadata_schema": "fastvideo.nanovideo_vsa_training", "metadata_schema_version": 2, "model": { "architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "kind": "wan-t2v-1.3b" }, "provenance": { "code": { "commit": "fefc437be38d0e1c3e231be5745b2d99af5446f6", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/home/hal-ywu/git/vsqa/outputs/training-source-fefc437-mixed600-c128-no-manifest-hash" } }, "configs": { "resolved": { "path": "resolved.yaml", "sha256": "7047bb89d1c6f40920c0f438f59ff775efe7a8c175212d0ff79c92a379d5af2d" }, "source": { "path": "source.yaml", "sha256": "dfeda6ad9b95acea42d430f87b7017a8daf60489471284ddd9ede39aaf4b14e7" } }, "lineage": { "checkpoint_id": "ckpt_fc684f3fae7a4148811693902e00dec7", "checkpoint_step": 500, "compression_gate_initialization": "inherited", "init_from": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0352-weighted-reconode-c128-k3-b16-1000-260922-104437/export-step500", "kind": "sparse_qat_weight_initialization", "optimizer_state": "fresh", "parent_architecture": { "_class_name": "WanTransformer3DModel", "_diffusers_version": "0.33.0.dev0", "added_kv_proj_dim": null, "attention_head_dim": 128, "cross_attn_norm": true, "eps": 1e-06, "ffn_dim": 8960, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ], "qk_norm": "rms_norm_across_heads", "rope_max_seq_len": 1024, "text_dim": 4096 }, "parent_attention_backend": "VSA_QAT_TRAIN_C128", "parent_cube_shape": [ 4, 4, 8 ], "parent_metadata": { "attention": { "backend": "VSA_QAT_TRAIN_C128", "compression_gates": "trained", "cube_shape": [ 4, 4, 8 ], "kind": "vsa", "physical_block_tokens": 64, "precision": { "coarse": "BF16", "compute": "BF16", "fine": "NVFP4 fake-quant with BF16 matmul", "master_weights": "FP32", "pv_fp8": false, "quant_p": true, "quant_v": true, "quantization_aware": true, "quantize_coarse": false, "quantize_fine": true, "two_level_p": false, "use_global_sf_p": false, "use_global_sf_qkv": false }, "selection": { "effective_topk": "max(1, min(ceil((1-sparsity)*num_blocks), num_blocks))", "mode": "top-p", "sparsity": 0.9 }, "training_kernel_family": "nanovideo_triton_sparse_bf16_mm" }, "checkpoint": { "created_at": "2026-09-22T14:21:51.738754+00:00", "id": "ckpt_fc684f3fae7a4148811693902e00dec7", "state": "training", "step": 500 }, "checkpoint_state": "inference_export", "code_provenance": { "commit": "fefc437be38d0e1c3e231be5745b2d99af5446f6", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/home/hal-ywu/git/vsqa/outputs/training-source-fefc437-mixed600-c128-reconode" } }, "config": { "callbacks": { "grad_clip": { "_target_": "fastvideo.train.callbacks.grad_clip.GradNormClipCallback", "max_grad_norm": 10.0 }, "validation": { "_target_": "fastvideo.train.methods.reconstruction_ode.validation.ReconstructionODEValidationCallback", "attn_qat_infer": false, "dataset_file": "/home/hal-ywu/git/vsqa/outputs/ode-v2-k3-260920/validation_16.json", "every_steps": 50, "guidance_scale": 1.0, "height": 448, "num_frames": 61, "offload_training_state": true, "pipeline_target": "fastvideo.pipelines.basic.wan.wan_shifted_dmd_pipeline.WanShiftedDMDPipeline", "run_at_start": false, "sampling_steps": [ 3 ], "unload_pipeline_after_validation": true, "use_validation_media_conditioning": false, "width": 896 } }, "method": { "_target_": "fastvideo.train.methods.reconstruction_ode.method.ReconstructionODEMethod", "attention_weight": 0.25, "checkpoint_blocks": true, "loss": "mse", "mlp_weight": 0.25, "ode_weight": 0.5, "student_steps": 3, "trainable_scope": "quantized_path" }, "models": { "student": { "_target_": "fastvideo.train.methods.reconstruction_ode.model.ReconstructionODEWanModel", "attention_backend": "VSA_QAT_TRAIN_C128", "enable_gradient_checkpointing_type": null, "flow_shift": 8.0, "init_from": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538/export-step500", "trainable": true }, "teacher": { "_target_": "fastvideo.train.models.wan.WanModel", "attention_backend": "TORCH_SDPA", "disable_custom_init_weights": true, "enable_gradient_checkpointing_type": null, "flow_shift": 8.0, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "trainable": false } }, "pipeline": { "dit_config": { "quant_config": "nvfp4_qat_train" }, "dmd_denoising_steps": [ 1000.0, 941.1763916015625, 800.0 ], "flow_shift": 8.0, "warp_denoising_step": false }, "training": { "checkpoint": { "checkpoints_total_limit": 0, "output_dir": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0352-weighted-reconode-c128-k3-b16-1000-260922-104437/checkpoints", "resume_from_checkpoint": "", "training_state_checkpointing_steps": 50 }, "data": { "data_path": "/home/hal-ywu/git/vsqa/data/ode-init-mixed-v2-1k12-48s-61f", "dataloader_num_workers": 0, "num_frames": 61, "num_height": 448, "num_latent_t": 16, "num_width": 896, "seed": 1000, "train_batch_size": 4, "training_cfg_rate": 0.0 }, "distributed": { "hsdp_replicate_dim": 4, "hsdp_shard_dim": 1, "num_gpus": 4, "sp_size": 1, "tp_size": 1 }, "dit_precision": "fp32", "loop": { "gradient_accumulation_steps": 1, "max_train_steps": 1000 }, "model": { "precondition_outputs": false, "weighting_scheme": "uniform" }, "optimizer": { "betas": [ 0.9, 0.999 ], "learning_rate": 1e-05, "lr_scheduler": "constant", "lr_warmup_steps": 0, "weight_decay": 0.01 }, "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0352-weighted-reconode-c128-k3-b16-1000-260922-104437", "trackers": [ "wandb" ] }, "vsa": { "cache_tile_buf": false, "sparsity": 0.9 } } }, "config_sha256": { "resolved": "88503181405aa091c2b17216b67da7e5fb4df91a6a896613edb11c9ac726892c", "source": "56ca3fe42f9f3d7c103185840bebfc31322d098591d8a73ae65a4afcb2902183" }, "metadata_schema": "fastvideo.nanovideo_vsa_training", "metadata_schema_version": 2, "model": { "architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "kind": "wan-t2v-1.3b" }, "provenance": { "code": { "commit": "fefc437be38d0e1c3e231be5745b2d99af5446f6", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/home/hal-ywu/git/vsqa/outputs/training-source-fefc437-mixed600-c128-reconode" } }, "configs": { "resolved": { "path": "resolved.yaml", "sha256": "88503181405aa091c2b17216b67da7e5fb4df91a6a896613edb11c9ac726892c" }, "source": { "path": "source.yaml", "sha256": "56ca3fe42f9f3d7c103185840bebfc31322d098591d8a73ae65a4afcb2902183" } }, "lineage": { "checkpoint_id": "ckpt_bd46a3b64697495fa252c0d5d3132b44", "checkpoint_step": 500, "compression_gate_initialization": "zero", "init_from": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538/export-step500", "kind": "dense_qat_weight_initialization", "optimizer_state": "fresh", "parent_architecture": { "_class_name": "WanTransformer3DModel", "_diffusers_version": "0.33.0.dev0", "added_kv_proj_dim": null, "attention_head_dim": 128, "cross_attn_norm": true, "eps": 1e-06, "ffn_dim": 8960, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ], "qk_norm": "rms_norm_across_heads", "rope_max_seq_len": 1024, "text_dim": 4096 }, "parent_metadata": { "attention": { "cube_shape": null, "kind": "dense", "precision": { "backward": "consistent STE with auxiliary output", "compute": "BF16", "group_size": 16, "master_weights": "FP32", "pv_format": "NVFP4 E2M1", "pv_fp8": false, "pv_scale_dtype": "FP8 E4M3FN", "qk_format": "NVFP4 E2M1", "qk_scale_dtype": "FP8 E4M3FN", "quant_p": true, "quant_v": true, "quantization_aware": true, "sm100_exact_m": false, "sm100_forward_mode": "fast", "sm100_optimized": true, "sm120_join_qat_pv": true, "smooth_k": false, "smooth_q": false, "two_level_p": false, "use_global_sf_p": false, "use_global_sf_qkv": false }, "training_kernel_family": "fastvideo_triton_dense_nvfp4_fake" }, "checkpoint": { "created_at": "2026-09-21T10:54:31.387678+00:00", "id": "ckpt_bd46a3b64697495fa252c0d5d3132b44", "state": "training", "step": 500 }, "checkpoint_state": "training_dcp", "code_provenance": { "commit": "fefc437be38d0e1c3e231be5745b2d99af5446f6", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/home/hal-ywu/git/vsqa/outputs/training-source-fefc437-recon-g5-no-manifest-hash" } }, "config": { "callbacks": { "grad_clip": { "_target_": "fastvideo.train.callbacks.grad_clip.GradNormClipCallback", "max_grad_norm": 10.0 }, "validation": { "_target_": "fastvideo.train.methods.layer_reconstruction.validation.LayerReconstructionValidationCallback", "attn_qat_infer": false, "dataset_file": "/home/hal-ywu/git/vsqa/docs/experiments/E0029-fastvideo-mixkit-training-migration/validation_12.json", "every_steps": 50, "guidance_scale": 5.0, "height": 448, "num_frames": 61, "pipeline_target": "fastvideo.pipelines.basic.wan.wan_pipeline.WanPipeline", "run_at_start": false, "sampling_steps": [ 50 ], "width": 896 } }, "method": { "_target_": "fastvideo.train.methods.layer_reconstruction.method.LayerReconstructionMethod", "checkpoint_blocks": true, "energy_floor": 1e-06, "group_size": 5, "input_source": "ode_start", "loss": "mse", "propagation": "teacher", "trainable_scope": "quantized_path" }, "models": { "student": { "_target_": "fastvideo.train.methods.layer_reconstruction.model.LayerReconstructionWanModel", "attention_backend": "ATTN_QAT_TRAIN", "enable_gradient_checkpointing_type": null, "flow_shift": 8.0, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "trainable": true }, "teacher": { "_target_": "fastvideo.train.models.wan.WanModel", "attention_backend": "TORCH_SDPA", "disable_custom_init_weights": true, "enable_gradient_checkpointing_type": null, "flow_shift": 8.0, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "trainable": false } }, "pipeline": { "dit_config": { "quant_config": "nvfp4_qat_train" }, "flow_shift": 8.0 }, "training": { "checkpoint": { "checkpoints_total_limit": 0, "output_dir": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538/checkpoints", "training_state_checkpointing_steps": 50 }, "data": { "data_path": "/home/hal-ywu/git/vsqa/data/ode-init-mixed-v2-1k12-48s-61f", "dataloader_num_workers": 0, "num_frames": 61, "num_height": 448, "num_latent_t": 16, "num_width": 896, "seed": 1000, "train_batch_size": 4, "training_cfg_rate": 0.0 }, "distributed": { "hsdp_replicate_dim": 4, "hsdp_shard_dim": 1, "num_gpus": 4, "sp_size": 1, "tp_size": 1 }, "dit_precision": "fp32", "loop": { "gradient_accumulation_steps": 1, "max_train_steps": 1000 }, "model": { "precondition_outputs": false, "weighting_scheme": "uniform" }, "optimizer": { "betas": [ 0.9, 0.999 ], "learning_rate": 1e-06, "lr_scheduler": "constant", "lr_warmup_steps": 0, "weight_decay": 0.0 }, "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538", "trackers": [ "wandb" ] } } }, "config_sha256": { "resolved": "e85bdb109864dc0c6830d51f49c4620c070f3ce03938b12000c251b71d7063c5", "source": "960720e845fba98343853b301fe382a789680f31fac0a6319b593136c61851e9" }, "metadata_schema": "fastvideo.dense_nvfp4_training", "metadata_schema_version": 2, "model": { "architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "kind": "wan-t2v-1.3b" }, "provenance": { "code": { "commit": "fefc437be38d0e1c3e231be5745b2d99af5446f6", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/home/hal-ywu/git/vsqa/outputs/training-source-fefc437-recon-g5-no-manifest-hash" } }, "configs": { "resolved": { "path": "resolved.yaml", "sha256": "e85bdb109864dc0c6830d51f49c4620c070f3ce03938b12000c251b71d7063c5" }, "source": { "path": "source.yaml", "sha256": "960720e845fba98343853b301fe382a789680f31fac0a6319b593136c61851e9" } }, "lineage": { "compression_gate_initialization": null, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "kind": "base_model_initialization", "optimizer_state": "fresh" } }, "run_identity": { "output_dir": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538/checkpoints", "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538", "trackers": [ "wandb" ] }, "wandb": { "id": "bwm9ne7h", "url": "https://wandb.ai/alexzms-ucsd/fastvideo/runs/bwm9ne7h" } }, "step": 500, "training_contract": { "attention_backend": "ATTN_QAT_TRAIN", "attention_kind": "dense", "cube_shape": null, "direct_lineage": { "compression_gate_initialization": null, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "kind": "base_model_initialization", "optimizer_state": "fresh" }, "kernel_source_files": { "fastvideo-kernel/python/fastvideo_kernel/triton_kernels/attn_qat_train.py": "09d90a4e3bae1a98caedd36ab71f8c7f72d896a8edc62e1e47e32440e645fa4c", "fastvideo-kernel/python/fastvideo_kernel/triton_kernels/quant_utils.py": "9bf123b175e11d667740182dc88712d2785a8798f1dd0228311d0f833c331c1f", "fastvideo/attention/backends/attn_qat_train.py": "c5853c4680269df5ccaa325c166ad881346f627248ad5793080899a2da905d14" }, "linear_quantization": { "activation_global_scale": "dynamic, 448*6/max|X| recomputed every forward", "backward": "whole-linear STE on the saved unquantized X and master W in the gradient dtype", "format": "NVFP4 W4A4: E2M1 values, 16-element blocks, FP8 E4M3 block scales, FP32 global scale", "forward": "flashinfer nvfp4_quantize + mm_fp4 (cutlass) real FP4 GEMM, BF16 output", "kind": "nvfp4_qat_train", "master_weights": "FP32", "scope": "student transformer only; teacher and critic load with quant_config dropped", "target_prefix_substrings": [ "ffn.fc_in", "ffn.fc_out", "to_q", "to_k", "to_v", "to_out", "self_attention.out_layer", "cross_attention.out_layer", "feed_forward.mlp.fc_in", "feed_forward.mlp.fc_out" ], "weight_global_scale": "dynamic, 448*6/max|W| recomputed every forward from the master weight" }, "model_architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "model_kind": "Wan bidirectional T2V", "precision": { "backward": "consistent STE with auxiliary output", "compute": "BF16", "group_size": 16, "master_weights": "FP32", "pv_format": "NVFP4 E2M1", "pv_fp8": false, "pv_scale_dtype": "FP8 E4M3FN", "qk_format": "NVFP4 E2M1", "qk_scale_dtype": "FP8 E4M3FN", "quant_p": true, "quant_v": true, "quantization_aware": true, "sm100_exact_m": false, "sm100_forward_mode": "fast", "sm100_optimized": true, "sm120_join_qat_pv": true, "smooth_k": false, "smooth_q": false, "two_level_p": false, "use_global_sf_p": false, "use_global_sf_qkv": false }, "selection": null, "supervision": { "cache": { "conditioning": "teacher_cfg", "geometry": { "num_frames": 61, "num_height": 448, "num_latent_t": 16, "num_width": 896 }, "pair_sampling": "all_twelve_teacher_anchors_to_clean", "pairs": 12000, "runtime": { "capability": [ 10, 0 ], "cuda": "12.8", "device_type": "cuda", "torch": "2.9.1+cu128" }, "schedule": { "flow_shift": 8.0, "sigmas": [ 1.0, 0.9973474740982056, 0.9945946335792542, 0.9917355179786682, 0.9887640476226807, 0.9856733679771423, 0.9824561476707458, 0.9791044592857361, 0.9756097793579102, 0.9719626307487488, 0.9681528210639954, 0.9641693830490112, 0.9599999785423279, 0.9556313753128052, 0.9510489702224731, 0.9462365508079529, 0.9411764144897461, 0.9358490705490112, 0.930232584476471, 0.9243027567863464, 0.9180328249931335, 0.9113923907279968, 0.9043477773666382, 0.8968610167503357, 0.8888888955116272, 0.880382776260376, 0.8712871074676514, 0.8615384697914124, 0.8510638475418091, 0.8397790193557739, 0.8275862336158752, 0.8143712878227234, 0.800000011920929, 0.7843137383460999, 0.7671233415603638, 0.7482014298439026, 0.7272727489471436, 0.7039999961853027, 0.6779661178588867, 0.6486486196517944, 0.6153846383094788, 0.5773195624351501, 0.5333333611488342, 0.4819277226924896, 0.42105263471603394, 0.3478260934352875, 0.25806450843811035, 0.145454540848732, 0.0 ], "solver": "shifted_euler_bf16_state", "student_indices": [ 0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44 ], "teacher_steps": 48, "timestep_parameterization": "sigma_times_1000", "timesteps": [ 1000.0, 997.3474731445312, 994.5946044921875, 991.7355346679688, 988.7640380859375, 985.67333984375, 982.4561767578125, 979.1044311523438, 975.6098022460938, 971.962646484375, 968.15283203125, 964.1693725585938, 960.0, 955.63134765625, 951.0489501953125, 946.236572265625, 941.1763916015625, 935.8490600585938, 930.2326049804688, 924.302734375, 918.0328369140625, 911.3923950195312, 904.3477783203125, 896.8610229492188, 888.888916015625, 880.3827514648438, 871.287109375, 861.5384521484375, 851.0638427734375, 839.7789916992188, 827.5862426757812, 814.3712768554688, 800.0, 784.313720703125, 767.1233520507812, 748.201416015625, 727.2727661132812, 704.0, 677.9661254882812, 648.6486206054688, 615.3846435546875, 577.319580078125, 533.3333740234375, 481.927734375, 421.0526428222656, 347.8260803222656, 258.06451416015625, 145.4545440673828 ] }, "source_kind": "mixed", "teacher": { "architecture": { "_class_name": "WanTransformer3DModel", "_diffusers_version": "0.33.0.dev0", "added_kv_proj_dim": null, "attention_head_dim": 128, "cross_attn_norm": true, "eps": 1e-06, "ffn_dim": 8960, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ], "qk_norm": "rms_norm_across_heads", "rope_max_seq_len": 1024, "text_dim": 4096 }, "backend": "TORCH_SDPA", "compute_dtype": "bfloat16", "files": { "transformer/config.json": "0b093fa072e9ff28763febe9b964ee582f566733a6d6709deb9dfba1bde16b81", "transformer/diffusion_pytorch_model-00001-of-00002.safetensors": "6d011927dbd2cc8afe53d57abab04a8fd86f615d83324770d985fb058ece3a24", "transformer/diffusion_pytorch_model-00002-of-00002.safetensors": "b92ec2309b1f239af6f746431815a881afcc938abb26a4f08d9a2fd6c892f872", "vae/config.json": "f0c1cc1d7decb5badc384f54691746a27a9aeff49f7ebca974e583389342d527" }, "guidance_scale": 6.0, "model_kind": "wan-t2v-1.3b", "negative_prompt": "\u8272\u8c03\u8273\u4e3d\uff0c\u8fc7\u66dd\uff0c\u9759\u6001\uff0c\u7ec6\u8282\u6a21\u7cca\u4e0d\u6e05\uff0c\u5b57\u5e55\uff0c\u98ce\u683c\uff0c\u4f5c\u54c1\uff0c\u753b\u4f5c\uff0c\u753b\u9762\uff0c\u9759\u6b62\uff0c\u6574\u4f53\u53d1\u7070\uff0c\u6700\u5dee\u8d28\u91cf\uff0c\u4f4e\u8d28\u91cf\uff0cJPEG\u538b\u7f29\u6b8b\u7559\uff0c\u4e11\u964b\u7684\uff0c\u6b8b\u7f3a\u7684\uff0c\u591a\u4f59\u7684\u624b\u6307\uff0c\u753b\u5f97\u4e0d\u597d\u7684\u624b\u90e8\uff0c\u753b\u5f97\u4e0d\u597d\u7684\u8138\u90e8\uff0c\u7578\u5f62\u7684\uff0c\u6bc1\u5bb9\u7684\uff0c\u5f62\u6001\u7578\u5f62\u7684\u80a2\u4f53\uff0c\u624b\u6307\u878d\u5408\uff0c\u9759\u6b62\u4e0d\u52a8\u7684\u753b\u9762\uff0c\u6742\u4e71\u7684\u80cc\u666f\uff0c\u4e09\u6761\u817f\uff0c\u80cc\u666f\u4eba\u5f88\u591a\uff0c\u5012\u7740\u8d70", "state_dtype": "bfloat16" }, "trajectories": 1000 }, "checkpoint_blocks": true, "endpoint_loss": false, "energy_floor": 1e-06, "forward_path": "teacher_trajectory_input_at_every_group_start_with_in_group_student_propagation_and_identical_frozen_embedding_and_conditioning", "gradient_scope": "group_local_student_quantized_path_groups_of_5", "group_size": 5, "input_description": "cached_ode_start_latent_and_corresponding_timestep_sigma_without_renoising", "input_source": "ode_start", "loss": "mse", "normalized_mse_denominator": "detached_per_sample_reference_block_update_mean_square_clamped_to_energy_floor", "objective": "fixed_teacher_layer_reconstruction", "propagation": "teacher", "protocol_version": 2, "reduction": "float32_per_sample_token_channel_mean_then_batch_and_layer_mean", "reference": { "attention_backend": "TORCH_SDPA", "frozen": true, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "linear_quantization": null, "weights_sha256": "aac925e262c5c2266456830496021a66598b03253c8e05b6fa048b1744e0f7a1" }, "target": "teacher_trajectory_hidden_state_after_each_group_of_blocks", "trainable_modules": [ "to_q", "to_k", "to_v", "to_out", "norm_q", "norm_k", "attn2", "ffn" ], "trainable_scope": "quantized_path" }, "training_kernel_family": "fastvideo_triton_dense_nvfp4_fake" } } } }, "run_identity": { "output_dir": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0352-weighted-reconode-c128-k3-b16-1000-260922-104437/checkpoints", "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0352-weighted-reconode-c128-k3-b16-1000-260922-104437", "trackers": [ "wandb" ] }, "wandb": { "id": "txtm6fh0", "url": "https://wandb.ai/alexzms-ucsd/fastvideo/runs/txtm6fh0" } }, "step": 500, "training_contract": { "attention_backend": "VSA_QAT_TRAIN_C128", "attention_kind": "VSA", "cube_shape": [ 4, 4, 8 ], "direct_lineage": { "checkpoint_id": "ckpt_bd46a3b64697495fa252c0d5d3132b44", "checkpoint_step": 500, "compression_gate_initialization": "zero", "init_from": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538/export-step500", "kind": "dense_qat_weight_initialization", "optimizer_state": "fresh", "parent_architecture": { "_class_name": "WanTransformer3DModel", "_diffusers_version": "0.33.0.dev0", "added_kv_proj_dim": null, "attention_head_dim": 128, "cross_attn_norm": true, "eps": 1e-06, "ffn_dim": 8960, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ], "qk_norm": "rms_norm_across_heads", "rope_max_seq_len": 1024, "text_dim": 4096 }, "parent_metadata": { "attention": { "cube_shape": null, "kind": "dense", "precision": { "backward": "consistent STE with auxiliary output", "compute": "BF16", "group_size": 16, "master_weights": "FP32", "pv_format": "NVFP4 E2M1", "pv_fp8": false, "pv_scale_dtype": "FP8 E4M3FN", "qk_format": "NVFP4 E2M1", "qk_scale_dtype": "FP8 E4M3FN", "quant_p": true, "quant_v": true, "quantization_aware": true, "sm100_exact_m": false, "sm100_forward_mode": "fast", "sm100_optimized": true, "sm120_join_qat_pv": true, "smooth_k": false, "smooth_q": false, "two_level_p": false, "use_global_sf_p": false, "use_global_sf_qkv": false }, "training_kernel_family": "fastvideo_triton_dense_nvfp4_fake" }, "checkpoint": { "created_at": "2026-09-21T10:54:31.387678+00:00", "id": "ckpt_bd46a3b64697495fa252c0d5d3132b44", "state": "training", "step": 500 }, "checkpoint_state": "training_dcp", "code_provenance": { "commit": "fefc437be38d0e1c3e231be5745b2d99af5446f6", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/home/hal-ywu/git/vsqa/outputs/training-source-fefc437-recon-g5-no-manifest-hash" } }, "config": { "callbacks": { "grad_clip": { "_target_": "fastvideo.train.callbacks.grad_clip.GradNormClipCallback", "max_grad_norm": 10.0 }, "validation": { "_target_": "fastvideo.train.methods.layer_reconstruction.validation.LayerReconstructionValidationCallback", "attn_qat_infer": false, "dataset_file": "/home/hal-ywu/git/vsqa/docs/experiments/E0029-fastvideo-mixkit-training-migration/validation_12.json", "every_steps": 50, "guidance_scale": 5.0, "height": 448, "num_frames": 61, "pipeline_target": "fastvideo.pipelines.basic.wan.wan_pipeline.WanPipeline", "run_at_start": false, "sampling_steps": [ 50 ], "width": 896 } }, "method": { "_target_": "fastvideo.train.methods.layer_reconstruction.method.LayerReconstructionMethod", "checkpoint_blocks": true, "energy_floor": 1e-06, "group_size": 5, "input_source": "ode_start", "loss": "mse", "propagation": "teacher", "trainable_scope": "quantized_path" }, "models": { "student": { "_target_": "fastvideo.train.methods.layer_reconstruction.model.LayerReconstructionWanModel", "attention_backend": "ATTN_QAT_TRAIN", "enable_gradient_checkpointing_type": null, "flow_shift": 8.0, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "trainable": true }, "teacher": { "_target_": "fastvideo.train.models.wan.WanModel", "attention_backend": "TORCH_SDPA", "disable_custom_init_weights": true, "enable_gradient_checkpointing_type": null, "flow_shift": 8.0, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "trainable": false } }, "pipeline": { "dit_config": { "quant_config": "nvfp4_qat_train" }, "flow_shift": 8.0 }, "training": { "checkpoint": { "checkpoints_total_limit": 0, "output_dir": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538/checkpoints", "training_state_checkpointing_steps": 50 }, "data": { "data_path": "/home/hal-ywu/git/vsqa/data/ode-init-mixed-v2-1k12-48s-61f", "dataloader_num_workers": 0, "num_frames": 61, "num_height": 448, "num_latent_t": 16, "num_width": 896, "seed": 1000, "train_batch_size": 4, "training_cfg_rate": 0.0 }, "distributed": { "hsdp_replicate_dim": 4, "hsdp_shard_dim": 1, "num_gpus": 4, "sp_size": 1, "tp_size": 1 }, "dit_precision": "fp32", "loop": { "gradient_accumulation_steps": 1, "max_train_steps": 1000 }, "model": { "precondition_outputs": false, "weighting_scheme": "uniform" }, "optimizer": { "betas": [ 0.9, 0.999 ], "learning_rate": 1e-06, "lr_scheduler": "constant", "lr_warmup_steps": 0, "weight_decay": 0.0 }, "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538", "trackers": [ "wandb" ] } } }, "config_sha256": { "resolved": "e85bdb109864dc0c6830d51f49c4620c070f3ce03938b12000c251b71d7063c5", "source": "960720e845fba98343853b301fe382a789680f31fac0a6319b593136c61851e9" }, "metadata_schema": "fastvideo.dense_nvfp4_training", "metadata_schema_version": 2, "model": { "architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "kind": "wan-t2v-1.3b" }, "provenance": { "code": { "commit": "fefc437be38d0e1c3e231be5745b2d99af5446f6", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/home/hal-ywu/git/vsqa/outputs/training-source-fefc437-recon-g5-no-manifest-hash" } }, "configs": { "resolved": { "path": "resolved.yaml", "sha256": "e85bdb109864dc0c6830d51f49c4620c070f3ce03938b12000c251b71d7063c5" }, "source": { "path": "source.yaml", "sha256": "960720e845fba98343853b301fe382a789680f31fac0a6319b593136c61851e9" } }, "lineage": { "compression_gate_initialization": null, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "kind": "base_model_initialization", "optimizer_state": "fresh" } }, "run_identity": { "output_dir": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538/checkpoints", "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538", "trackers": [ "wandb" ] }, "wandb": { "id": "bwm9ne7h", "url": "https://wandb.ai/alexzms-ucsd/fastvideo/runs/bwm9ne7h" } }, "step": 500, "training_contract": { "attention_backend": "ATTN_QAT_TRAIN", "attention_kind": "dense", "cube_shape": null, "direct_lineage": { "compression_gate_initialization": null, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "kind": "base_model_initialization", "optimizer_state": "fresh" }, "kernel_source_files": { "fastvideo-kernel/python/fastvideo_kernel/triton_kernels/attn_qat_train.py": "09d90a4e3bae1a98caedd36ab71f8c7f72d896a8edc62e1e47e32440e645fa4c", "fastvideo-kernel/python/fastvideo_kernel/triton_kernels/quant_utils.py": "9bf123b175e11d667740182dc88712d2785a8798f1dd0228311d0f833c331c1f", "fastvideo/attention/backends/attn_qat_train.py": "c5853c4680269df5ccaa325c166ad881346f627248ad5793080899a2da905d14" }, "linear_quantization": { "activation_global_scale": "dynamic, 448*6/max|X| recomputed every forward", "backward": "whole-linear STE on the saved unquantized X and master W in the gradient dtype", "format": "NVFP4 W4A4: E2M1 values, 16-element blocks, FP8 E4M3 block scales, FP32 global scale", "forward": "flashinfer nvfp4_quantize + mm_fp4 (cutlass) real FP4 GEMM, BF16 output", "kind": "nvfp4_qat_train", "master_weights": "FP32", "scope": "student transformer only; teacher and critic load with quant_config dropped", "target_prefix_substrings": [ "ffn.fc_in", "ffn.fc_out", "to_q", "to_k", "to_v", "to_out", "self_attention.out_layer", "cross_attention.out_layer", "feed_forward.mlp.fc_in", "feed_forward.mlp.fc_out" ], "weight_global_scale": "dynamic, 448*6/max|W| recomputed every forward from the master weight" }, "model_architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "model_kind": "Wan bidirectional T2V", "precision": { "backward": "consistent STE with auxiliary output", "compute": "BF16", "group_size": 16, "master_weights": "FP32", "pv_format": "NVFP4 E2M1", "pv_fp8": false, "pv_scale_dtype": "FP8 E4M3FN", "qk_format": "NVFP4 E2M1", "qk_scale_dtype": "FP8 E4M3FN", "quant_p": true, "quant_v": true, "quantization_aware": true, "sm100_exact_m": false, "sm100_forward_mode": "fast", "sm100_optimized": true, "sm120_join_qat_pv": true, "smooth_k": false, "smooth_q": false, "two_level_p": false, "use_global_sf_p": false, "use_global_sf_qkv": false }, "selection": null, "supervision": { "cache": { "conditioning": "teacher_cfg", "geometry": { "num_frames": 61, "num_height": 448, "num_latent_t": 16, "num_width": 896 }, "pair_sampling": "all_twelve_teacher_anchors_to_clean", "pairs": 12000, "runtime": { "capability": [ 10, 0 ], "cuda": "12.8", "device_type": "cuda", "torch": "2.9.1+cu128" }, "schedule": { "flow_shift": 8.0, "sigmas": [ 1.0, 0.9973474740982056, 0.9945946335792542, 0.9917355179786682, 0.9887640476226807, 0.9856733679771423, 0.9824561476707458, 0.9791044592857361, 0.9756097793579102, 0.9719626307487488, 0.9681528210639954, 0.9641693830490112, 0.9599999785423279, 0.9556313753128052, 0.9510489702224731, 0.9462365508079529, 0.9411764144897461, 0.9358490705490112, 0.930232584476471, 0.9243027567863464, 0.9180328249931335, 0.9113923907279968, 0.9043477773666382, 0.8968610167503357, 0.8888888955116272, 0.880382776260376, 0.8712871074676514, 0.8615384697914124, 0.8510638475418091, 0.8397790193557739, 0.8275862336158752, 0.8143712878227234, 0.800000011920929, 0.7843137383460999, 0.7671233415603638, 0.7482014298439026, 0.7272727489471436, 0.7039999961853027, 0.6779661178588867, 0.6486486196517944, 0.6153846383094788, 0.5773195624351501, 0.5333333611488342, 0.4819277226924896, 0.42105263471603394, 0.3478260934352875, 0.25806450843811035, 0.145454540848732, 0.0 ], "solver": "shifted_euler_bf16_state", "student_indices": [ 0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44 ], "teacher_steps": 48, "timestep_parameterization": "sigma_times_1000", "timesteps": [ 1000.0, 997.3474731445312, 994.5946044921875, 991.7355346679688, 988.7640380859375, 985.67333984375, 982.4561767578125, 979.1044311523438, 975.6098022460938, 971.962646484375, 968.15283203125, 964.1693725585938, 960.0, 955.63134765625, 951.0489501953125, 946.236572265625, 941.1763916015625, 935.8490600585938, 930.2326049804688, 924.302734375, 918.0328369140625, 911.3923950195312, 904.3477783203125, 896.8610229492188, 888.888916015625, 880.3827514648438, 871.287109375, 861.5384521484375, 851.0638427734375, 839.7789916992188, 827.5862426757812, 814.3712768554688, 800.0, 784.313720703125, 767.1233520507812, 748.201416015625, 727.2727661132812, 704.0, 677.9661254882812, 648.6486206054688, 615.3846435546875, 577.319580078125, 533.3333740234375, 481.927734375, 421.0526428222656, 347.8260803222656, 258.06451416015625, 145.4545440673828 ] }, "source_kind": "mixed", "teacher": { "architecture": { "_class_name": "WanTransformer3DModel", "_diffusers_version": "0.33.0.dev0", "added_kv_proj_dim": null, "attention_head_dim": 128, "cross_attn_norm": true, "eps": 1e-06, "ffn_dim": 8960, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ], "qk_norm": "rms_norm_across_heads", "rope_max_seq_len": 1024, "text_dim": 4096 }, "backend": "TORCH_SDPA", "compute_dtype": "bfloat16", "files": { "transformer/config.json": "0b093fa072e9ff28763febe9b964ee582f566733a6d6709deb9dfba1bde16b81", "transformer/diffusion_pytorch_model-00001-of-00002.safetensors": "6d011927dbd2cc8afe53d57abab04a8fd86f615d83324770d985fb058ece3a24", "transformer/diffusion_pytorch_model-00002-of-00002.safetensors": "b92ec2309b1f239af6f746431815a881afcc938abb26a4f08d9a2fd6c892f872", "vae/config.json": "f0c1cc1d7decb5badc384f54691746a27a9aeff49f7ebca974e583389342d527" }, "guidance_scale": 6.0, "model_kind": "wan-t2v-1.3b", "negative_prompt": "\u8272\u8c03\u8273\u4e3d\uff0c\u8fc7\u66dd\uff0c\u9759\u6001\uff0c\u7ec6\u8282\u6a21\u7cca\u4e0d\u6e05\uff0c\u5b57\u5e55\uff0c\u98ce\u683c\uff0c\u4f5c\u54c1\uff0c\u753b\u4f5c\uff0c\u753b\u9762\uff0c\u9759\u6b62\uff0c\u6574\u4f53\u53d1\u7070\uff0c\u6700\u5dee\u8d28\u91cf\uff0c\u4f4e\u8d28\u91cf\uff0cJPEG\u538b\u7f29\u6b8b\u7559\uff0c\u4e11\u964b\u7684\uff0c\u6b8b\u7f3a\u7684\uff0c\u591a\u4f59\u7684\u624b\u6307\uff0c\u753b\u5f97\u4e0d\u597d\u7684\u624b\u90e8\uff0c\u753b\u5f97\u4e0d\u597d\u7684\u8138\u90e8\uff0c\u7578\u5f62\u7684\uff0c\u6bc1\u5bb9\u7684\uff0c\u5f62\u6001\u7578\u5f62\u7684\u80a2\u4f53\uff0c\u624b\u6307\u878d\u5408\uff0c\u9759\u6b62\u4e0d\u52a8\u7684\u753b\u9762\uff0c\u6742\u4e71\u7684\u80cc\u666f\uff0c\u4e09\u6761\u817f\uff0c\u80cc\u666f\u4eba\u5f88\u591a\uff0c\u5012\u7740\u8d70", "state_dtype": "bfloat16" }, "trajectories": 1000 }, "checkpoint_blocks": true, "endpoint_loss": false, "energy_floor": 1e-06, "forward_path": "teacher_trajectory_input_at_every_group_start_with_in_group_student_propagation_and_identical_frozen_embedding_and_conditioning", "gradient_scope": "group_local_student_quantized_path_groups_of_5", "group_size": 5, "input_description": "cached_ode_start_latent_and_corresponding_timestep_sigma_without_renoising", "input_source": "ode_start", "loss": "mse", "normalized_mse_denominator": "detached_per_sample_reference_block_update_mean_square_clamped_to_energy_floor", "objective": "fixed_teacher_layer_reconstruction", "propagation": "teacher", "protocol_version": 2, "reduction": "float32_per_sample_token_channel_mean_then_batch_and_layer_mean", "reference": { "attention_backend": "TORCH_SDPA", "frozen": true, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "linear_quantization": null, "weights_sha256": "aac925e262c5c2266456830496021a66598b03253c8e05b6fa048b1744e0f7a1" }, "target": "teacher_trajectory_hidden_state_after_each_group_of_blocks", "trainable_modules": [ "to_q", "to_k", "to_v", "to_out", "norm_q", "norm_k", "attn2", "ffn" ], "trainable_scope": "quantized_path" }, "training_kernel_family": "fastvideo_triton_dense_nvfp4_fake" } } }, "kernel_source_commit": "e8b4630e0ee15be13d1e0f6823882b776edbe62e", "kernel_source_files": { "block_sparse_attn.py": "da77938f5e68822e57b7571507064199c62440d580efa9e3b14f9b5d1bc72199", "qat/nvfp4_utils.py": "70b613fca22817ca764496095efc73127515574a53e4b9fcc70bd56a63b3ac3e", "qat/quant_utils.py": "e450efca2c375d48889c9d647f5760c9fe90330436dfc623a5684ebe6a48080d", "qat/sparse/bf16_mm.py": "1c8bbf7ac573ccfa3d28b41d001c922cb261c1c74bcc8e83743e0da092c82ff5", "triton_kernels/block_sparse_attn.py": "715d4d51ff26f3edf9764b6708769376068f228bedd1f7db26aee93375203d83", "triton_kernels/index.py": "ce8beccf1f66554917765976820834fdf03a45ad0d1852e6da0b16c811d7b19d" }, "linear_quantization": { "activation_global_scale": "dynamic, 448*6/max|X| recomputed every forward", "backward": "whole-linear STE on the saved unquantized X and master W in the gradient dtype", "format": "NVFP4 W4A4: E2M1 values, 16-element blocks, FP8 E4M3 block scales, FP32 global scale", "forward": "flashinfer nvfp4_quantize + mm_fp4 (cutlass) real FP4 GEMM, BF16 output", "kind": "nvfp4_qat_train", "master_weights": "FP32", "scope": "student transformer only; the frozen BF16 teacher loads with quant_config dropped", "target_prefix_substrings": [ "ffn.fc_in", "ffn.fc_out", "to_q", "to_k", "to_v", "to_out", "self_attention.out_layer", "cross_attention.out_layer", "feed_forward.mlp.fc_in", "feed_forward.mlp.fc_out" ], "weight_global_scale": "dynamic, 448*6/max|W| recomputed every forward from the master weight" }, "model_architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "model_kind": "Wan bidirectional T2V", "physical_block_tokens": 64, "precision": { "coarse": "BF16", "compute": "BF16", "fine": "NVFP4 fake-quant with BF16 matmul", "master_weights": "FP32", "pv_fp8": false, "quant_p": true, "quant_v": true, "quantize_coarse": false, "quantize_fine": true, "two_level_p": false, "use_global_sf_p": false, "use_global_sf_qkv": false }, "selection": { "effective_topk": "max(1, min(ceil((1-sparsity)*num_blocks), num_blocks))", "mode": "top-p", "sparsity": 0.9 }, "supervision": { "attention_backend": "VSA_QAT_TRAIN_C128", "attention_kind": "vsa", "attention_metadata": "student_vsa_metadata_at_recipe_sparsity_for_both_forwards", "attention_weight": 0.25, "cache": { "conditioning": "teacher_cfg", "geometry": { "num_frames": 61, "num_height": 448, "num_latent_t": 16, "num_width": 896 }, "pair_sampling": "all_twelve_teacher_anchors_to_clean", "pairs": 12000, "runtime": { "capability": [ 10, 0 ], "cuda": "12.8", "device_type": "cuda", "torch": "2.9.1+cu128" }, "schedule": { "flow_shift": 8.0, "sigmas": [ 1.0, 0.9973474740982056, 0.9945946335792542, 0.9917355179786682, 0.9887640476226807, 0.9856733679771423, 0.9824561476707458, 0.9791044592857361, 0.9756097793579102, 0.9719626307487488, 0.9681528210639954, 0.9641693830490112, 0.9599999785423279, 0.9556313753128052, 0.9510489702224731, 0.9462365508079529, 0.9411764144897461, 0.9358490705490112, 0.930232584476471, 0.9243027567863464, 0.9180328249931335, 0.9113923907279968, 0.9043477773666382, 0.8968610167503357, 0.8888888955116272, 0.880382776260376, 0.8712871074676514, 0.8615384697914124, 0.8510638475418091, 0.8397790193557739, 0.8275862336158752, 0.8143712878227234, 0.800000011920929, 0.7843137383460999, 0.7671233415603638, 0.7482014298439026, 0.7272727489471436, 0.7039999961853027, 0.6779661178588867, 0.6486486196517944, 0.6153846383094788, 0.5773195624351501, 0.5333333611488342, 0.4819277226924896, 0.42105263471603394, 0.3478260934352875, 0.25806450843811035, 0.145454540848732, 0.0 ], "solver": "shifted_euler_bf16_state", "student_indices": [ 0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44 ], "teacher_steps": 48, "timestep_parameterization": "sigma_times_1000", "timesteps": [ 1000.0, 997.3474731445312, 994.5946044921875, 991.7355346679688, 988.7640380859375, 985.67333984375, 982.4561767578125, 979.1044311523438, 975.6098022460938, 971.962646484375, 968.15283203125, 964.1693725585938, 960.0, 955.63134765625, 951.0489501953125, 946.236572265625, 941.1763916015625, 935.8490600585938, 930.2326049804688, 924.302734375, 918.0328369140625, 911.3923950195312, 904.3477783203125, 896.8610229492188, 888.888916015625, 880.3827514648438, 871.287109375, 861.5384521484375, 851.0638427734375, 839.7789916992188, 827.5862426757812, 814.3712768554688, 800.0, 784.313720703125, 767.1233520507812, 748.201416015625, 727.2727661132812, 704.0, 677.9661254882812, 648.6486206054688, 615.3846435546875, 577.319580078125, 533.3333740234375, 481.927734375, 421.0526428222656, 347.8260803222656, 258.06451416015625, 145.4545440673828 ] }, "source_kind": "mixed", "teacher": { "architecture": { "_class_name": "WanTransformer3DModel", "_diffusers_version": "0.33.0.dev0", "added_kv_proj_dim": null, "attention_head_dim": 128, "cross_attn_norm": true, "eps": 1e-06, "ffn_dim": 8960, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ], "qk_norm": "rms_norm_across_heads", "rope_max_seq_len": 1024, "text_dim": 4096 }, "backend": "TORCH_SDPA", "compute_dtype": "bfloat16", "files": { "transformer/config.json": "0b093fa072e9ff28763febe9b964ee582f566733a6d6709deb9dfba1bde16b81", "transformer/diffusion_pytorch_model-00001-of-00002.safetensors": "6d011927dbd2cc8afe53d57abab04a8fd86f615d83324770d985fb058ece3a24", "transformer/diffusion_pytorch_model-00002-of-00002.safetensors": "b92ec2309b1f239af6f746431815a881afcc938abb26a4f08d9a2fd6c892f872", "vae/config.json": "f0c1cc1d7decb5badc384f54691746a27a9aeff49f7ebca974e583389342d527" }, "guidance_scale": 6.0, "model_kind": "wan-t2v-1.3b", "negative_prompt": "\u8272\u8c03\u8273\u4e3d\uff0c\u8fc7\u66dd\uff0c\u9759\u6001\uff0c\u7ec6\u8282\u6a21\u7cca\u4e0d\u6e05\uff0c\u5b57\u5e55\uff0c\u98ce\u683c\uff0c\u4f5c\u54c1\uff0c\u753b\u4f5c\uff0c\u753b\u9762\uff0c\u9759\u6b62\uff0c\u6574\u4f53\u53d1\u7070\uff0c\u6700\u5dee\u8d28\u91cf\uff0c\u4f4e\u8d28\u91cf\uff0cJPEG\u538b\u7f29\u6b8b\u7559\uff0c\u4e11\u964b\u7684\uff0c\u6b8b\u7f3a\u7684\uff0c\u591a\u4f59\u7684\u624b\u6307\uff0c\u753b\u5f97\u4e0d\u597d\u7684\u624b\u90e8\uff0c\u753b\u5f97\u4e0d\u597d\u7684\u8138\u90e8\uff0c\u7578\u5f62\u7684\uff0c\u6bc1\u5bb9\u7684\uff0c\u5f62\u6001\u7578\u5f62\u7684\u80a2\u4f53\uff0c\u624b\u6307\u878d\u5408\uff0c\u9759\u6b62\u4e0d\u52a8\u7684\u753b\u9762\uff0c\u6742\u4e71\u7684\u80cc\u666f\uff0c\u4e09\u6761\u817f\uff0c\u80cc\u666f\u4eba\u5f88\u591a\uff0c\u5012\u7740\u8d70", "state_dtype": "bfloat16" }, "trajectories": 1000 }, "checkpoint_blocks": true, "cube_shape": [ 4, 4, 8 ], "endpoint_formula": "mean((start_latent - start_sigma * student_velocity - target_latent)^2)", "endpoint_loss": true, "energy_floor": 1e-06, "forward_path": "teacher_driven_local_reconstruction_plus_separate_full_student_velocity_forward_both_under_student_sparse_block_selection", "gradient_scope": "local_reconstruction_plus_end_to_end_ode_within_student_quantized_path", "input_source": "cached_ode_start_latent_and_corresponding_timestep_sigma_without_renoising", "loss": "mse", "mlp_weight": 0.25, "objective": "attention_mlp_reconstruction_and_ode_endpoint", "ode_weight": 0.5, "protocol_version": 2, "reconstruction_normalization": "none", "reduction": "attention_and_mlp_each_layer_mean_plus_ode_mean_mse_weighted_sum", "reference": { "attention_backend": "TORCH_SDPA", "attention_kind": "dense", "frozen": true, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "linear_quantization": null, "weights_sha256": "aac925e262c5c2266456830496021a66598b03253c8e05b6fa048b1744e0f7a1" }, "student_initialization": { "checkpoint_step": 500, "compression_gate_initialization": "zero", "init_from": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538/export-step500", "initial_weight_equality_with_reference": false, "kind": "dense_qat_weight_initialization" }, "student_steps": 3, "target": [ "self_attention_post_output_projection_pre_gate_residual", "ffn_pre_gate_residual", "cached_teacher_final_clean_latent" ], "trainable_modules": [ "to_q", "to_k", "to_v", "to_out", "norm_q", "norm_k", "attn2", "ffn", "to_gate_compress" ], "trainable_scope": "quantized_path", "training_schedule": { "flow_shift": 8.0, "indices": [ 0, 16, 32 ], "sigmas": [ 1.0, 0.9411764144897461, 0.800000011920929 ], "student_steps": 3, "teacher_steps": 48, "timesteps": [ 1000.0, 941.1763916015625, 800.0 ], "version": 1 } }, "training_kernel_family": "nanovideo_triton_sparse_bf16_mm" } }, "parent_quantization_aware": true } }, "run_identity": { "output_dir": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0354-dmd-c128-k3-b16-1000-260922-104437/checkpoints", "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0354-dmd-c128-k3-b16-1000-260922-104437", "trackers": [ "wandb" ] }, "wandb": { "id": "z9z855ib", "url": "https://wandb.ai/alexzms-ucsd/fastvideo/runs/z9z855ib" } }, "step": 1000, "training_contract": { "attention_backend": "VSA_QAT_TRAIN_C128", "attention_kind": "VSA", "cube_shape": [ 4, 4, 8 ], "direct_lineage": { "checkpoint_id": "ckpt_fc684f3fae7a4148811693902e00dec7", "checkpoint_step": 500, "compression_gate_initialization": "inherited", "init_from": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0352-weighted-reconode-c128-k3-b16-1000-260922-104437/export-step500", "kind": "sparse_qat_weight_initialization", "optimizer_state": "fresh", "parent_architecture": { "_class_name": "WanTransformer3DModel", "_diffusers_version": "0.33.0.dev0", "added_kv_proj_dim": null, "attention_head_dim": 128, "cross_attn_norm": true, "eps": 1e-06, "ffn_dim": 8960, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ], "qk_norm": "rms_norm_across_heads", "rope_max_seq_len": 1024, "text_dim": 4096 }, "parent_attention_backend": "VSA_QAT_TRAIN_C128", "parent_cube_shape": [ 4, 4, 8 ], "parent_metadata": { "attention": { "backend": "VSA_QAT_TRAIN_C128", "compression_gates": "trained", "cube_shape": [ 4, 4, 8 ], "kind": "vsa", "physical_block_tokens": 64, "precision": { "coarse": "BF16", "compute": "BF16", "fine": "NVFP4 fake-quant with BF16 matmul", "master_weights": "FP32", "pv_fp8": false, "quant_p": true, "quant_v": true, "quantization_aware": true, "quantize_coarse": false, "quantize_fine": true, "two_level_p": false, "use_global_sf_p": false, "use_global_sf_qkv": false }, "selection": { "effective_topk": "max(1, min(ceil((1-sparsity)*num_blocks), num_blocks))", "mode": "top-p", "sparsity": 0.9 }, "training_kernel_family": "nanovideo_triton_sparse_bf16_mm" }, "checkpoint": { "created_at": "2026-09-22T14:21:51.738754+00:00", "id": "ckpt_fc684f3fae7a4148811693902e00dec7", "state": "training", "step": 500 }, "checkpoint_state": "inference_export", "code_provenance": { "commit": "fefc437be38d0e1c3e231be5745b2d99af5446f6", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/home/hal-ywu/git/vsqa/outputs/training-source-fefc437-mixed600-c128-reconode" } }, "config": { "callbacks": { "grad_clip": { "_target_": "fastvideo.train.callbacks.grad_clip.GradNormClipCallback", "max_grad_norm": 10.0 }, "validation": { "_target_": "fastvideo.train.methods.reconstruction_ode.validation.ReconstructionODEValidationCallback", "attn_qat_infer": false, "dataset_file": "/home/hal-ywu/git/vsqa/outputs/ode-v2-k3-260920/validation_16.json", "every_steps": 50, "guidance_scale": 1.0, "height": 448, "num_frames": 61, "offload_training_state": true, "pipeline_target": "fastvideo.pipelines.basic.wan.wan_shifted_dmd_pipeline.WanShiftedDMDPipeline", "run_at_start": false, "sampling_steps": [ 3 ], "unload_pipeline_after_validation": true, "use_validation_media_conditioning": false, "width": 896 } }, "method": { "_target_": "fastvideo.train.methods.reconstruction_ode.method.ReconstructionODEMethod", "attention_weight": 0.25, "checkpoint_blocks": true, "loss": "mse", "mlp_weight": 0.25, "ode_weight": 0.5, "student_steps": 3, "trainable_scope": "quantized_path" }, "models": { "student": { "_target_": "fastvideo.train.methods.reconstruction_ode.model.ReconstructionODEWanModel", "attention_backend": "VSA_QAT_TRAIN_C128", "enable_gradient_checkpointing_type": null, "flow_shift": 8.0, "init_from": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538/export-step500", "trainable": true }, "teacher": { "_target_": "fastvideo.train.models.wan.WanModel", "attention_backend": "TORCH_SDPA", "disable_custom_init_weights": true, "enable_gradient_checkpointing_type": null, "flow_shift": 8.0, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "trainable": false } }, "pipeline": { "dit_config": { "quant_config": "nvfp4_qat_train" }, "dmd_denoising_steps": [ 1000.0, 941.1763916015625, 800.0 ], "flow_shift": 8.0, "warp_denoising_step": false }, "training": { "checkpoint": { "checkpoints_total_limit": 0, "output_dir": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0352-weighted-reconode-c128-k3-b16-1000-260922-104437/checkpoints", "resume_from_checkpoint": "", "training_state_checkpointing_steps": 50 }, "data": { "data_path": "/home/hal-ywu/git/vsqa/data/ode-init-mixed-v2-1k12-48s-61f", "dataloader_num_workers": 0, "num_frames": 61, "num_height": 448, "num_latent_t": 16, "num_width": 896, "seed": 1000, "train_batch_size": 4, "training_cfg_rate": 0.0 }, "distributed": { "hsdp_replicate_dim": 4, "hsdp_shard_dim": 1, "num_gpus": 4, "sp_size": 1, "tp_size": 1 }, "dit_precision": "fp32", "loop": { "gradient_accumulation_steps": 1, "max_train_steps": 1000 }, "model": { "precondition_outputs": false, "weighting_scheme": "uniform" }, "optimizer": { "betas": [ 0.9, 0.999 ], "learning_rate": 1e-05, "lr_scheduler": "constant", "lr_warmup_steps": 0, "weight_decay": 0.01 }, "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0352-weighted-reconode-c128-k3-b16-1000-260922-104437", "trackers": [ "wandb" ] }, "vsa": { "cache_tile_buf": false, "sparsity": 0.9 } } }, "config_sha256": { "resolved": "88503181405aa091c2b17216b67da7e5fb4df91a6a896613edb11c9ac726892c", "source": "56ca3fe42f9f3d7c103185840bebfc31322d098591d8a73ae65a4afcb2902183" }, "metadata_schema": "fastvideo.nanovideo_vsa_training", "metadata_schema_version": 2, "model": { "architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "kind": "wan-t2v-1.3b" }, "provenance": { "code": { "commit": "fefc437be38d0e1c3e231be5745b2d99af5446f6", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/home/hal-ywu/git/vsqa/outputs/training-source-fefc437-mixed600-c128-reconode" } }, "configs": { "resolved": { "path": "resolved.yaml", "sha256": "88503181405aa091c2b17216b67da7e5fb4df91a6a896613edb11c9ac726892c" }, "source": { "path": "source.yaml", "sha256": "56ca3fe42f9f3d7c103185840bebfc31322d098591d8a73ae65a4afcb2902183" } }, "lineage": { "checkpoint_id": "ckpt_bd46a3b64697495fa252c0d5d3132b44", "checkpoint_step": 500, "compression_gate_initialization": "zero", "init_from": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538/export-step500", "kind": "dense_qat_weight_initialization", "optimizer_state": "fresh", "parent_architecture": { "_class_name": "WanTransformer3DModel", "_diffusers_version": "0.33.0.dev0", "added_kv_proj_dim": null, "attention_head_dim": 128, "cross_attn_norm": true, "eps": 1e-06, "ffn_dim": 8960, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ], "qk_norm": "rms_norm_across_heads", "rope_max_seq_len": 1024, "text_dim": 4096 }, "parent_metadata": { "attention": { "cube_shape": null, "kind": "dense", "precision": { "backward": "consistent STE with auxiliary output", "compute": "BF16", "group_size": 16, "master_weights": "FP32", "pv_format": "NVFP4 E2M1", "pv_fp8": false, "pv_scale_dtype": "FP8 E4M3FN", "qk_format": "NVFP4 E2M1", "qk_scale_dtype": "FP8 E4M3FN", "quant_p": true, "quant_v": true, "quantization_aware": true, "sm100_exact_m": false, "sm100_forward_mode": "fast", "sm100_optimized": true, "sm120_join_qat_pv": true, "smooth_k": false, "smooth_q": false, "two_level_p": false, "use_global_sf_p": false, "use_global_sf_qkv": false }, "training_kernel_family": "fastvideo_triton_dense_nvfp4_fake" }, "checkpoint": { "created_at": "2026-09-21T10:54:31.387678+00:00", "id": "ckpt_bd46a3b64697495fa252c0d5d3132b44", "state": "training", "step": 500 }, "checkpoint_state": "training_dcp", "code_provenance": { "commit": "fefc437be38d0e1c3e231be5745b2d99af5446f6", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/home/hal-ywu/git/vsqa/outputs/training-source-fefc437-recon-g5-no-manifest-hash" } }, "config": { "callbacks": { "grad_clip": { "_target_": "fastvideo.train.callbacks.grad_clip.GradNormClipCallback", "max_grad_norm": 10.0 }, "validation": { "_target_": "fastvideo.train.methods.layer_reconstruction.validation.LayerReconstructionValidationCallback", "attn_qat_infer": false, "dataset_file": "/home/hal-ywu/git/vsqa/docs/experiments/E0029-fastvideo-mixkit-training-migration/validation_12.json", "every_steps": 50, "guidance_scale": 5.0, "height": 448, "num_frames": 61, "pipeline_target": "fastvideo.pipelines.basic.wan.wan_pipeline.WanPipeline", "run_at_start": false, "sampling_steps": [ 50 ], "width": 896 } }, "method": { "_target_": "fastvideo.train.methods.layer_reconstruction.method.LayerReconstructionMethod", "checkpoint_blocks": true, "energy_floor": 1e-06, "group_size": 5, "input_source": "ode_start", "loss": "mse", "propagation": "teacher", "trainable_scope": "quantized_path" }, "models": { "student": { "_target_": "fastvideo.train.methods.layer_reconstruction.model.LayerReconstructionWanModel", "attention_backend": "ATTN_QAT_TRAIN", "enable_gradient_checkpointing_type": null, "flow_shift": 8.0, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "trainable": true }, "teacher": { "_target_": "fastvideo.train.models.wan.WanModel", "attention_backend": "TORCH_SDPA", "disable_custom_init_weights": true, "enable_gradient_checkpointing_type": null, "flow_shift": 8.0, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "trainable": false } }, "pipeline": { "dit_config": { "quant_config": "nvfp4_qat_train" }, "flow_shift": 8.0 }, "training": { "checkpoint": { "checkpoints_total_limit": 0, "output_dir": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538/checkpoints", "training_state_checkpointing_steps": 50 }, "data": { "data_path": "/home/hal-ywu/git/vsqa/data/ode-init-mixed-v2-1k12-48s-61f", "dataloader_num_workers": 0, "num_frames": 61, "num_height": 448, "num_latent_t": 16, "num_width": 896, "seed": 1000, "train_batch_size": 4, "training_cfg_rate": 0.0 }, "distributed": { "hsdp_replicate_dim": 4, "hsdp_shard_dim": 1, "num_gpus": 4, "sp_size": 1, "tp_size": 1 }, "dit_precision": "fp32", "loop": { "gradient_accumulation_steps": 1, "max_train_steps": 1000 }, "model": { "precondition_outputs": false, "weighting_scheme": "uniform" }, "optimizer": { "betas": [ 0.9, 0.999 ], "learning_rate": 1e-06, "lr_scheduler": "constant", "lr_warmup_steps": 0, "weight_decay": 0.0 }, "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538", "trackers": [ "wandb" ] } } }, "config_sha256": { "resolved": "e85bdb109864dc0c6830d51f49c4620c070f3ce03938b12000c251b71d7063c5", "source": "960720e845fba98343853b301fe382a789680f31fac0a6319b593136c61851e9" }, "metadata_schema": "fastvideo.dense_nvfp4_training", "metadata_schema_version": 2, "model": { "architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "kind": "wan-t2v-1.3b" }, "provenance": { "code": { "commit": "fefc437be38d0e1c3e231be5745b2d99af5446f6", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/home/hal-ywu/git/vsqa/outputs/training-source-fefc437-recon-g5-no-manifest-hash" } }, "configs": { "resolved": { "path": "resolved.yaml", "sha256": "e85bdb109864dc0c6830d51f49c4620c070f3ce03938b12000c251b71d7063c5" }, "source": { "path": "source.yaml", "sha256": "960720e845fba98343853b301fe382a789680f31fac0a6319b593136c61851e9" } }, "lineage": { "compression_gate_initialization": null, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "kind": "base_model_initialization", "optimizer_state": "fresh" } }, "run_identity": { "output_dir": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538/checkpoints", "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538", "trackers": [ "wandb" ] }, "wandb": { "id": "bwm9ne7h", "url": "https://wandb.ai/alexzms-ucsd/fastvideo/runs/bwm9ne7h" } }, "step": 500, "training_contract": { "attention_backend": "ATTN_QAT_TRAIN", "attention_kind": "dense", "cube_shape": null, "direct_lineage": { "compression_gate_initialization": null, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "kind": "base_model_initialization", "optimizer_state": "fresh" }, "kernel_source_files": { "fastvideo-kernel/python/fastvideo_kernel/triton_kernels/attn_qat_train.py": "09d90a4e3bae1a98caedd36ab71f8c7f72d896a8edc62e1e47e32440e645fa4c", "fastvideo-kernel/python/fastvideo_kernel/triton_kernels/quant_utils.py": "9bf123b175e11d667740182dc88712d2785a8798f1dd0228311d0f833c331c1f", "fastvideo/attention/backends/attn_qat_train.py": "c5853c4680269df5ccaa325c166ad881346f627248ad5793080899a2da905d14" }, "linear_quantization": { "activation_global_scale": "dynamic, 448*6/max|X| recomputed every forward", "backward": "whole-linear STE on the saved unquantized X and master W in the gradient dtype", "format": "NVFP4 W4A4: E2M1 values, 16-element blocks, FP8 E4M3 block scales, FP32 global scale", "forward": "flashinfer nvfp4_quantize + mm_fp4 (cutlass) real FP4 GEMM, BF16 output", "kind": "nvfp4_qat_train", "master_weights": "FP32", "scope": "student transformer only; teacher and critic load with quant_config dropped", "target_prefix_substrings": [ "ffn.fc_in", "ffn.fc_out", "to_q", "to_k", "to_v", "to_out", "self_attention.out_layer", "cross_attention.out_layer", "feed_forward.mlp.fc_in", "feed_forward.mlp.fc_out" ], "weight_global_scale": "dynamic, 448*6/max|W| recomputed every forward from the master weight" }, "model_architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "model_kind": "Wan bidirectional T2V", "precision": { "backward": "consistent STE with auxiliary output", "compute": "BF16", "group_size": 16, "master_weights": "FP32", "pv_format": "NVFP4 E2M1", "pv_fp8": false, "pv_scale_dtype": "FP8 E4M3FN", "qk_format": "NVFP4 E2M1", "qk_scale_dtype": "FP8 E4M3FN", "quant_p": true, "quant_v": true, "quantization_aware": true, "sm100_exact_m": false, "sm100_forward_mode": "fast", "sm100_optimized": true, "sm120_join_qat_pv": true, "smooth_k": false, "smooth_q": false, "two_level_p": false, "use_global_sf_p": false, "use_global_sf_qkv": false }, "selection": null, "supervision": { "cache": { "conditioning": "teacher_cfg", "geometry": { "num_frames": 61, "num_height": 448, "num_latent_t": 16, "num_width": 896 }, "pair_sampling": "all_twelve_teacher_anchors_to_clean", "pairs": 12000, "runtime": { "capability": [ 10, 0 ], "cuda": "12.8", "device_type": "cuda", "torch": "2.9.1+cu128" }, "schedule": { "flow_shift": 8.0, "sigmas": [ 1.0, 0.9973474740982056, 0.9945946335792542, 0.9917355179786682, 0.9887640476226807, 0.9856733679771423, 0.9824561476707458, 0.9791044592857361, 0.9756097793579102, 0.9719626307487488, 0.9681528210639954, 0.9641693830490112, 0.9599999785423279, 0.9556313753128052, 0.9510489702224731, 0.9462365508079529, 0.9411764144897461, 0.9358490705490112, 0.930232584476471, 0.9243027567863464, 0.9180328249931335, 0.9113923907279968, 0.9043477773666382, 0.8968610167503357, 0.8888888955116272, 0.880382776260376, 0.8712871074676514, 0.8615384697914124, 0.8510638475418091, 0.8397790193557739, 0.8275862336158752, 0.8143712878227234, 0.800000011920929, 0.7843137383460999, 0.7671233415603638, 0.7482014298439026, 0.7272727489471436, 0.7039999961853027, 0.6779661178588867, 0.6486486196517944, 0.6153846383094788, 0.5773195624351501, 0.5333333611488342, 0.4819277226924896, 0.42105263471603394, 0.3478260934352875, 0.25806450843811035, 0.145454540848732, 0.0 ], "solver": "shifted_euler_bf16_state", "student_indices": [ 0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44 ], "teacher_steps": 48, "timestep_parameterization": "sigma_times_1000", "timesteps": [ 1000.0, 997.3474731445312, 994.5946044921875, 991.7355346679688, 988.7640380859375, 985.67333984375, 982.4561767578125, 979.1044311523438, 975.6098022460938, 971.962646484375, 968.15283203125, 964.1693725585938, 960.0, 955.63134765625, 951.0489501953125, 946.236572265625, 941.1763916015625, 935.8490600585938, 930.2326049804688, 924.302734375, 918.0328369140625, 911.3923950195312, 904.3477783203125, 896.8610229492188, 888.888916015625, 880.3827514648438, 871.287109375, 861.5384521484375, 851.0638427734375, 839.7789916992188, 827.5862426757812, 814.3712768554688, 800.0, 784.313720703125, 767.1233520507812, 748.201416015625, 727.2727661132812, 704.0, 677.9661254882812, 648.6486206054688, 615.3846435546875, 577.319580078125, 533.3333740234375, 481.927734375, 421.0526428222656, 347.8260803222656, 258.06451416015625, 145.4545440673828 ] }, "source_kind": "mixed", "teacher": { "architecture": { "_class_name": "WanTransformer3DModel", "_diffusers_version": "0.33.0.dev0", "added_kv_proj_dim": null, "attention_head_dim": 128, "cross_attn_norm": true, "eps": 1e-06, "ffn_dim": 8960, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ], "qk_norm": "rms_norm_across_heads", "rope_max_seq_len": 1024, "text_dim": 4096 }, "backend": "TORCH_SDPA", "compute_dtype": "bfloat16", "files": { "transformer/config.json": "0b093fa072e9ff28763febe9b964ee582f566733a6d6709deb9dfba1bde16b81", "transformer/diffusion_pytorch_model-00001-of-00002.safetensors": "6d011927dbd2cc8afe53d57abab04a8fd86f615d83324770d985fb058ece3a24", "transformer/diffusion_pytorch_model-00002-of-00002.safetensors": "b92ec2309b1f239af6f746431815a881afcc938abb26a4f08d9a2fd6c892f872", "vae/config.json": "f0c1cc1d7decb5badc384f54691746a27a9aeff49f7ebca974e583389342d527" }, "guidance_scale": 6.0, "model_kind": "wan-t2v-1.3b", "negative_prompt": "\u8272\u8c03\u8273\u4e3d\uff0c\u8fc7\u66dd\uff0c\u9759\u6001\uff0c\u7ec6\u8282\u6a21\u7cca\u4e0d\u6e05\uff0c\u5b57\u5e55\uff0c\u98ce\u683c\uff0c\u4f5c\u54c1\uff0c\u753b\u4f5c\uff0c\u753b\u9762\uff0c\u9759\u6b62\uff0c\u6574\u4f53\u53d1\u7070\uff0c\u6700\u5dee\u8d28\u91cf\uff0c\u4f4e\u8d28\u91cf\uff0cJPEG\u538b\u7f29\u6b8b\u7559\uff0c\u4e11\u964b\u7684\uff0c\u6b8b\u7f3a\u7684\uff0c\u591a\u4f59\u7684\u624b\u6307\uff0c\u753b\u5f97\u4e0d\u597d\u7684\u624b\u90e8\uff0c\u753b\u5f97\u4e0d\u597d\u7684\u8138\u90e8\uff0c\u7578\u5f62\u7684\uff0c\u6bc1\u5bb9\u7684\uff0c\u5f62\u6001\u7578\u5f62\u7684\u80a2\u4f53\uff0c\u624b\u6307\u878d\u5408\uff0c\u9759\u6b62\u4e0d\u52a8\u7684\u753b\u9762\uff0c\u6742\u4e71\u7684\u80cc\u666f\uff0c\u4e09\u6761\u817f\uff0c\u80cc\u666f\u4eba\u5f88\u591a\uff0c\u5012\u7740\u8d70", "state_dtype": "bfloat16" }, "trajectories": 1000 }, "checkpoint_blocks": true, "endpoint_loss": false, "energy_floor": 1e-06, "forward_path": "teacher_trajectory_input_at_every_group_start_with_in_group_student_propagation_and_identical_frozen_embedding_and_conditioning", "gradient_scope": "group_local_student_quantized_path_groups_of_5", "group_size": 5, "input_description": "cached_ode_start_latent_and_corresponding_timestep_sigma_without_renoising", "input_source": "ode_start", "loss": "mse", "normalized_mse_denominator": "detached_per_sample_reference_block_update_mean_square_clamped_to_energy_floor", "objective": "fixed_teacher_layer_reconstruction", "propagation": "teacher", "protocol_version": 2, "reduction": "float32_per_sample_token_channel_mean_then_batch_and_layer_mean", "reference": { "attention_backend": "TORCH_SDPA", "frozen": true, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "linear_quantization": null, "weights_sha256": "aac925e262c5c2266456830496021a66598b03253c8e05b6fa048b1744e0f7a1" }, "target": "teacher_trajectory_hidden_state_after_each_group_of_blocks", "trainable_modules": [ "to_q", "to_k", "to_v", "to_out", "norm_q", "norm_k", "attn2", "ffn" ], "trainable_scope": "quantized_path" }, "training_kernel_family": "fastvideo_triton_dense_nvfp4_fake" } } } }, "run_identity": { "output_dir": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0352-weighted-reconode-c128-k3-b16-1000-260922-104437/checkpoints", "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0352-weighted-reconode-c128-k3-b16-1000-260922-104437", "trackers": [ "wandb" ] }, "wandb": { "id": "txtm6fh0", "url": "https://wandb.ai/alexzms-ucsd/fastvideo/runs/txtm6fh0" } }, "step": 500, "training_contract": { "attention_backend": "VSA_QAT_TRAIN_C128", "attention_kind": "VSA", "cube_shape": [ 4, 4, 8 ], "direct_lineage": { "checkpoint_id": "ckpt_bd46a3b64697495fa252c0d5d3132b44", "checkpoint_step": 500, "compression_gate_initialization": "zero", "init_from": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538/export-step500", "kind": "dense_qat_weight_initialization", "optimizer_state": "fresh", "parent_architecture": { "_class_name": "WanTransformer3DModel", "_diffusers_version": "0.33.0.dev0", "added_kv_proj_dim": null, "attention_head_dim": 128, "cross_attn_norm": true, "eps": 1e-06, "ffn_dim": 8960, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ], "qk_norm": "rms_norm_across_heads", "rope_max_seq_len": 1024, "text_dim": 4096 }, "parent_metadata": { "attention": { "cube_shape": null, "kind": "dense", "precision": { "backward": "consistent STE with auxiliary output", "compute": "BF16", "group_size": 16, "master_weights": "FP32", "pv_format": "NVFP4 E2M1", "pv_fp8": false, "pv_scale_dtype": "FP8 E4M3FN", "qk_format": "NVFP4 E2M1", "qk_scale_dtype": "FP8 E4M3FN", "quant_p": true, "quant_v": true, "quantization_aware": true, "sm100_exact_m": false, "sm100_forward_mode": "fast", "sm100_optimized": true, "sm120_join_qat_pv": true, "smooth_k": false, "smooth_q": false, "two_level_p": false, "use_global_sf_p": false, "use_global_sf_qkv": false }, "training_kernel_family": "fastvideo_triton_dense_nvfp4_fake" }, "checkpoint": { "created_at": "2026-09-21T10:54:31.387678+00:00", "id": "ckpt_bd46a3b64697495fa252c0d5d3132b44", "state": "training", "step": 500 }, "checkpoint_state": "training_dcp", "code_provenance": { "commit": "fefc437be38d0e1c3e231be5745b2d99af5446f6", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/home/hal-ywu/git/vsqa/outputs/training-source-fefc437-recon-g5-no-manifest-hash" } }, "config": { "callbacks": { "grad_clip": { "_target_": "fastvideo.train.callbacks.grad_clip.GradNormClipCallback", "max_grad_norm": 10.0 }, "validation": { "_target_": "fastvideo.train.methods.layer_reconstruction.validation.LayerReconstructionValidationCallback", "attn_qat_infer": false, "dataset_file": "/home/hal-ywu/git/vsqa/docs/experiments/E0029-fastvideo-mixkit-training-migration/validation_12.json", "every_steps": 50, "guidance_scale": 5.0, "height": 448, "num_frames": 61, "pipeline_target": "fastvideo.pipelines.basic.wan.wan_pipeline.WanPipeline", "run_at_start": false, "sampling_steps": [ 50 ], "width": 896 } }, "method": { "_target_": "fastvideo.train.methods.layer_reconstruction.method.LayerReconstructionMethod", "checkpoint_blocks": true, "energy_floor": 1e-06, "group_size": 5, "input_source": "ode_start", "loss": "mse", "propagation": "teacher", "trainable_scope": "quantized_path" }, "models": { "student": { "_target_": "fastvideo.train.methods.layer_reconstruction.model.LayerReconstructionWanModel", "attention_backend": "ATTN_QAT_TRAIN", "enable_gradient_checkpointing_type": null, "flow_shift": 8.0, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "trainable": true }, "teacher": { "_target_": "fastvideo.train.models.wan.WanModel", "attention_backend": "TORCH_SDPA", "disable_custom_init_weights": true, "enable_gradient_checkpointing_type": null, "flow_shift": 8.0, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "trainable": false } }, "pipeline": { "dit_config": { "quant_config": "nvfp4_qat_train" }, "flow_shift": 8.0 }, "training": { "checkpoint": { "checkpoints_total_limit": 0, "output_dir": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538/checkpoints", "training_state_checkpointing_steps": 50 }, "data": { "data_path": "/home/hal-ywu/git/vsqa/data/ode-init-mixed-v2-1k12-48s-61f", "dataloader_num_workers": 0, "num_frames": 61, "num_height": 448, "num_latent_t": 16, "num_width": 896, "seed": 1000, "train_batch_size": 4, "training_cfg_rate": 0.0 }, "distributed": { "hsdp_replicate_dim": 4, "hsdp_shard_dim": 1, "num_gpus": 4, "sp_size": 1, "tp_size": 1 }, "dit_precision": "fp32", "loop": { "gradient_accumulation_steps": 1, "max_train_steps": 1000 }, "model": { "precondition_outputs": false, "weighting_scheme": "uniform" }, "optimizer": { "betas": [ 0.9, 0.999 ], "learning_rate": 1e-06, "lr_scheduler": "constant", "lr_warmup_steps": 0, "weight_decay": 0.0 }, "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538", "trackers": [ "wandb" ] } } }, "config_sha256": { "resolved": "e85bdb109864dc0c6830d51f49c4620c070f3ce03938b12000c251b71d7063c5", "source": "960720e845fba98343853b301fe382a789680f31fac0a6319b593136c61851e9" }, "metadata_schema": "fastvideo.dense_nvfp4_training", "metadata_schema_version": 2, "model": { "architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "kind": "wan-t2v-1.3b" }, "provenance": { "code": { "commit": "fefc437be38d0e1c3e231be5745b2d99af5446f6", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/home/hal-ywu/git/vsqa/outputs/training-source-fefc437-recon-g5-no-manifest-hash" } }, "configs": { "resolved": { "path": "resolved.yaml", "sha256": "e85bdb109864dc0c6830d51f49c4620c070f3ce03938b12000c251b71d7063c5" }, "source": { "path": "source.yaml", "sha256": "960720e845fba98343853b301fe382a789680f31fac0a6319b593136c61851e9" } }, "lineage": { "compression_gate_initialization": null, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "kind": "base_model_initialization", "optimizer_state": "fresh" } }, "run_identity": { "output_dir": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538/checkpoints", "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538", "trackers": [ "wandb" ] }, "wandb": { "id": "bwm9ne7h", "url": "https://wandb.ai/alexzms-ucsd/fastvideo/runs/bwm9ne7h" } }, "step": 500, "training_contract": { "attention_backend": "ATTN_QAT_TRAIN", "attention_kind": "dense", "cube_shape": null, "direct_lineage": { "compression_gate_initialization": null, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "kind": "base_model_initialization", "optimizer_state": "fresh" }, "kernel_source_files": { "fastvideo-kernel/python/fastvideo_kernel/triton_kernels/attn_qat_train.py": "09d90a4e3bae1a98caedd36ab71f8c7f72d896a8edc62e1e47e32440e645fa4c", "fastvideo-kernel/python/fastvideo_kernel/triton_kernels/quant_utils.py": "9bf123b175e11d667740182dc88712d2785a8798f1dd0228311d0f833c331c1f", "fastvideo/attention/backends/attn_qat_train.py": "c5853c4680269df5ccaa325c166ad881346f627248ad5793080899a2da905d14" }, "linear_quantization": { "activation_global_scale": "dynamic, 448*6/max|X| recomputed every forward", "backward": "whole-linear STE on the saved unquantized X and master W in the gradient dtype", "format": "NVFP4 W4A4: E2M1 values, 16-element blocks, FP8 E4M3 block scales, FP32 global scale", "forward": "flashinfer nvfp4_quantize + mm_fp4 (cutlass) real FP4 GEMM, BF16 output", "kind": "nvfp4_qat_train", "master_weights": "FP32", "scope": "student transformer only; teacher and critic load with quant_config dropped", "target_prefix_substrings": [ "ffn.fc_in", "ffn.fc_out", "to_q", "to_k", "to_v", "to_out", "self_attention.out_layer", "cross_attention.out_layer", "feed_forward.mlp.fc_in", "feed_forward.mlp.fc_out" ], "weight_global_scale": "dynamic, 448*6/max|W| recomputed every forward from the master weight" }, "model_architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "model_kind": "Wan bidirectional T2V", "precision": { "backward": "consistent STE with auxiliary output", "compute": "BF16", "group_size": 16, "master_weights": "FP32", "pv_format": "NVFP4 E2M1", "pv_fp8": false, "pv_scale_dtype": "FP8 E4M3FN", "qk_format": "NVFP4 E2M1", "qk_scale_dtype": "FP8 E4M3FN", "quant_p": true, "quant_v": true, "quantization_aware": true, "sm100_exact_m": false, "sm100_forward_mode": "fast", "sm100_optimized": true, "sm120_join_qat_pv": true, "smooth_k": false, "smooth_q": false, "two_level_p": false, "use_global_sf_p": false, "use_global_sf_qkv": false }, "selection": null, "supervision": { "cache": { "conditioning": "teacher_cfg", "geometry": { "num_frames": 61, "num_height": 448, "num_latent_t": 16, "num_width": 896 }, "pair_sampling": "all_twelve_teacher_anchors_to_clean", "pairs": 12000, "runtime": { "capability": [ 10, 0 ], "cuda": "12.8", "device_type": "cuda", "torch": "2.9.1+cu128" }, "schedule": { "flow_shift": 8.0, "sigmas": [ 1.0, 0.9973474740982056, 0.9945946335792542, 0.9917355179786682, 0.9887640476226807, 0.9856733679771423, 0.9824561476707458, 0.9791044592857361, 0.9756097793579102, 0.9719626307487488, 0.9681528210639954, 0.9641693830490112, 0.9599999785423279, 0.9556313753128052, 0.9510489702224731, 0.9462365508079529, 0.9411764144897461, 0.9358490705490112, 0.930232584476471, 0.9243027567863464, 0.9180328249931335, 0.9113923907279968, 0.9043477773666382, 0.8968610167503357, 0.8888888955116272, 0.880382776260376, 0.8712871074676514, 0.8615384697914124, 0.8510638475418091, 0.8397790193557739, 0.8275862336158752, 0.8143712878227234, 0.800000011920929, 0.7843137383460999, 0.7671233415603638, 0.7482014298439026, 0.7272727489471436, 0.7039999961853027, 0.6779661178588867, 0.6486486196517944, 0.6153846383094788, 0.5773195624351501, 0.5333333611488342, 0.4819277226924896, 0.42105263471603394, 0.3478260934352875, 0.25806450843811035, 0.145454540848732, 0.0 ], "solver": "shifted_euler_bf16_state", "student_indices": [ 0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44 ], "teacher_steps": 48, "timestep_parameterization": "sigma_times_1000", "timesteps": [ 1000.0, 997.3474731445312, 994.5946044921875, 991.7355346679688, 988.7640380859375, 985.67333984375, 982.4561767578125, 979.1044311523438, 975.6098022460938, 971.962646484375, 968.15283203125, 964.1693725585938, 960.0, 955.63134765625, 951.0489501953125, 946.236572265625, 941.1763916015625, 935.8490600585938, 930.2326049804688, 924.302734375, 918.0328369140625, 911.3923950195312, 904.3477783203125, 896.8610229492188, 888.888916015625, 880.3827514648438, 871.287109375, 861.5384521484375, 851.0638427734375, 839.7789916992188, 827.5862426757812, 814.3712768554688, 800.0, 784.313720703125, 767.1233520507812, 748.201416015625, 727.2727661132812, 704.0, 677.9661254882812, 648.6486206054688, 615.3846435546875, 577.319580078125, 533.3333740234375, 481.927734375, 421.0526428222656, 347.8260803222656, 258.06451416015625, 145.4545440673828 ] }, "source_kind": "mixed", "teacher": { "architecture": { "_class_name": "WanTransformer3DModel", "_diffusers_version": "0.33.0.dev0", "added_kv_proj_dim": null, "attention_head_dim": 128, "cross_attn_norm": true, "eps": 1e-06, "ffn_dim": 8960, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ], "qk_norm": "rms_norm_across_heads", "rope_max_seq_len": 1024, "text_dim": 4096 }, "backend": "TORCH_SDPA", "compute_dtype": "bfloat16", "files": { "transformer/config.json": "0b093fa072e9ff28763febe9b964ee582f566733a6d6709deb9dfba1bde16b81", "transformer/diffusion_pytorch_model-00001-of-00002.safetensors": "6d011927dbd2cc8afe53d57abab04a8fd86f615d83324770d985fb058ece3a24", "transformer/diffusion_pytorch_model-00002-of-00002.safetensors": "b92ec2309b1f239af6f746431815a881afcc938abb26a4f08d9a2fd6c892f872", "vae/config.json": "f0c1cc1d7decb5badc384f54691746a27a9aeff49f7ebca974e583389342d527" }, "guidance_scale": 6.0, "model_kind": "wan-t2v-1.3b", "negative_prompt": "\u8272\u8c03\u8273\u4e3d\uff0c\u8fc7\u66dd\uff0c\u9759\u6001\uff0c\u7ec6\u8282\u6a21\u7cca\u4e0d\u6e05\uff0c\u5b57\u5e55\uff0c\u98ce\u683c\uff0c\u4f5c\u54c1\uff0c\u753b\u4f5c\uff0c\u753b\u9762\uff0c\u9759\u6b62\uff0c\u6574\u4f53\u53d1\u7070\uff0c\u6700\u5dee\u8d28\u91cf\uff0c\u4f4e\u8d28\u91cf\uff0cJPEG\u538b\u7f29\u6b8b\u7559\uff0c\u4e11\u964b\u7684\uff0c\u6b8b\u7f3a\u7684\uff0c\u591a\u4f59\u7684\u624b\u6307\uff0c\u753b\u5f97\u4e0d\u597d\u7684\u624b\u90e8\uff0c\u753b\u5f97\u4e0d\u597d\u7684\u8138\u90e8\uff0c\u7578\u5f62\u7684\uff0c\u6bc1\u5bb9\u7684\uff0c\u5f62\u6001\u7578\u5f62\u7684\u80a2\u4f53\uff0c\u624b\u6307\u878d\u5408\uff0c\u9759\u6b62\u4e0d\u52a8\u7684\u753b\u9762\uff0c\u6742\u4e71\u7684\u80cc\u666f\uff0c\u4e09\u6761\u817f\uff0c\u80cc\u666f\u4eba\u5f88\u591a\uff0c\u5012\u7740\u8d70", "state_dtype": "bfloat16" }, "trajectories": 1000 }, "checkpoint_blocks": true, "endpoint_loss": false, "energy_floor": 1e-06, "forward_path": "teacher_trajectory_input_at_every_group_start_with_in_group_student_propagation_and_identical_frozen_embedding_and_conditioning", "gradient_scope": "group_local_student_quantized_path_groups_of_5", "group_size": 5, "input_description": "cached_ode_start_latent_and_corresponding_timestep_sigma_without_renoising", "input_source": "ode_start", "loss": "mse", "normalized_mse_denominator": "detached_per_sample_reference_block_update_mean_square_clamped_to_energy_floor", "objective": "fixed_teacher_layer_reconstruction", "propagation": "teacher", "protocol_version": 2, "reduction": "float32_per_sample_token_channel_mean_then_batch_and_layer_mean", "reference": { "attention_backend": "TORCH_SDPA", "frozen": true, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "linear_quantization": null, "weights_sha256": "aac925e262c5c2266456830496021a66598b03253c8e05b6fa048b1744e0f7a1" }, "target": "teacher_trajectory_hidden_state_after_each_group_of_blocks", "trainable_modules": [ "to_q", "to_k", "to_v", "to_out", "norm_q", "norm_k", "attn2", "ffn" ], "trainable_scope": "quantized_path" }, "training_kernel_family": "fastvideo_triton_dense_nvfp4_fake" } } }, "kernel_source_commit": "e8b4630e0ee15be13d1e0f6823882b776edbe62e", "kernel_source_files": { "block_sparse_attn.py": "da77938f5e68822e57b7571507064199c62440d580efa9e3b14f9b5d1bc72199", "qat/nvfp4_utils.py": "70b613fca22817ca764496095efc73127515574a53e4b9fcc70bd56a63b3ac3e", "qat/quant_utils.py": "e450efca2c375d48889c9d647f5760c9fe90330436dfc623a5684ebe6a48080d", "qat/sparse/bf16_mm.py": "1c8bbf7ac573ccfa3d28b41d001c922cb261c1c74bcc8e83743e0da092c82ff5", "triton_kernels/block_sparse_attn.py": "715d4d51ff26f3edf9764b6708769376068f228bedd1f7db26aee93375203d83", "triton_kernels/index.py": "ce8beccf1f66554917765976820834fdf03a45ad0d1852e6da0b16c811d7b19d" }, "linear_quantization": { "activation_global_scale": "dynamic, 448*6/max|X| recomputed every forward", "backward": "whole-linear STE on the saved unquantized X and master W in the gradient dtype", "format": "NVFP4 W4A4: E2M1 values, 16-element blocks, FP8 E4M3 block scales, FP32 global scale", "forward": "flashinfer nvfp4_quantize + mm_fp4 (cutlass) real FP4 GEMM, BF16 output", "kind": "nvfp4_qat_train", "master_weights": "FP32", "scope": "student transformer only; the frozen BF16 teacher loads with quant_config dropped", "target_prefix_substrings": [ "ffn.fc_in", "ffn.fc_out", "to_q", "to_k", "to_v", "to_out", "self_attention.out_layer", "cross_attention.out_layer", "feed_forward.mlp.fc_in", "feed_forward.mlp.fc_out" ], "weight_global_scale": "dynamic, 448*6/max|W| recomputed every forward from the master weight" }, "model_architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "model_kind": "Wan bidirectional T2V", "physical_block_tokens": 64, "precision": { "coarse": "BF16", "compute": "BF16", "fine": "NVFP4 fake-quant with BF16 matmul", "master_weights": "FP32", "pv_fp8": false, "quant_p": true, "quant_v": true, "quantize_coarse": false, "quantize_fine": true, "two_level_p": false, "use_global_sf_p": false, "use_global_sf_qkv": false }, "selection": { "effective_topk": "max(1, min(ceil((1-sparsity)*num_blocks), num_blocks))", "mode": "top-p", "sparsity": 0.9 }, "supervision": { "attention_backend": "VSA_QAT_TRAIN_C128", "attention_kind": "vsa", "attention_metadata": "student_vsa_metadata_at_recipe_sparsity_for_both_forwards", "attention_weight": 0.25, "cache": { "conditioning": "teacher_cfg", "geometry": { "num_frames": 61, "num_height": 448, "num_latent_t": 16, "num_width": 896 }, "pair_sampling": "all_twelve_teacher_anchors_to_clean", "pairs": 12000, "runtime": { "capability": [ 10, 0 ], "cuda": "12.8", "device_type": "cuda", "torch": "2.9.1+cu128" }, "schedule": { "flow_shift": 8.0, "sigmas": [ 1.0, 0.9973474740982056, 0.9945946335792542, 0.9917355179786682, 0.9887640476226807, 0.9856733679771423, 0.9824561476707458, 0.9791044592857361, 0.9756097793579102, 0.9719626307487488, 0.9681528210639954, 0.9641693830490112, 0.9599999785423279, 0.9556313753128052, 0.9510489702224731, 0.9462365508079529, 0.9411764144897461, 0.9358490705490112, 0.930232584476471, 0.9243027567863464, 0.9180328249931335, 0.9113923907279968, 0.9043477773666382, 0.8968610167503357, 0.8888888955116272, 0.880382776260376, 0.8712871074676514, 0.8615384697914124, 0.8510638475418091, 0.8397790193557739, 0.8275862336158752, 0.8143712878227234, 0.800000011920929, 0.7843137383460999, 0.7671233415603638, 0.7482014298439026, 0.7272727489471436, 0.7039999961853027, 0.6779661178588867, 0.6486486196517944, 0.6153846383094788, 0.5773195624351501, 0.5333333611488342, 0.4819277226924896, 0.42105263471603394, 0.3478260934352875, 0.25806450843811035, 0.145454540848732, 0.0 ], "solver": "shifted_euler_bf16_state", "student_indices": [ 0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44 ], "teacher_steps": 48, "timestep_parameterization": "sigma_times_1000", "timesteps": [ 1000.0, 997.3474731445312, 994.5946044921875, 991.7355346679688, 988.7640380859375, 985.67333984375, 982.4561767578125, 979.1044311523438, 975.6098022460938, 971.962646484375, 968.15283203125, 964.1693725585938, 960.0, 955.63134765625, 951.0489501953125, 946.236572265625, 941.1763916015625, 935.8490600585938, 930.2326049804688, 924.302734375, 918.0328369140625, 911.3923950195312, 904.3477783203125, 896.8610229492188, 888.888916015625, 880.3827514648438, 871.287109375, 861.5384521484375, 851.0638427734375, 839.7789916992188, 827.5862426757812, 814.3712768554688, 800.0, 784.313720703125, 767.1233520507812, 748.201416015625, 727.2727661132812, 704.0, 677.9661254882812, 648.6486206054688, 615.3846435546875, 577.319580078125, 533.3333740234375, 481.927734375, 421.0526428222656, 347.8260803222656, 258.06451416015625, 145.4545440673828 ] }, "source_kind": "mixed", "teacher": { "architecture": { "_class_name": "WanTransformer3DModel", "_diffusers_version": "0.33.0.dev0", "added_kv_proj_dim": null, "attention_head_dim": 128, "cross_attn_norm": true, "eps": 1e-06, "ffn_dim": 8960, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ], "qk_norm": "rms_norm_across_heads", "rope_max_seq_len": 1024, "text_dim": 4096 }, "backend": "TORCH_SDPA", "compute_dtype": "bfloat16", "files": { "transformer/config.json": "0b093fa072e9ff28763febe9b964ee582f566733a6d6709deb9dfba1bde16b81", "transformer/diffusion_pytorch_model-00001-of-00002.safetensors": "6d011927dbd2cc8afe53d57abab04a8fd86f615d83324770d985fb058ece3a24", "transformer/diffusion_pytorch_model-00002-of-00002.safetensors": "b92ec2309b1f239af6f746431815a881afcc938abb26a4f08d9a2fd6c892f872", "vae/config.json": "f0c1cc1d7decb5badc384f54691746a27a9aeff49f7ebca974e583389342d527" }, "guidance_scale": 6.0, "model_kind": "wan-t2v-1.3b", "negative_prompt": "\u8272\u8c03\u8273\u4e3d\uff0c\u8fc7\u66dd\uff0c\u9759\u6001\uff0c\u7ec6\u8282\u6a21\u7cca\u4e0d\u6e05\uff0c\u5b57\u5e55\uff0c\u98ce\u683c\uff0c\u4f5c\u54c1\uff0c\u753b\u4f5c\uff0c\u753b\u9762\uff0c\u9759\u6b62\uff0c\u6574\u4f53\u53d1\u7070\uff0c\u6700\u5dee\u8d28\u91cf\uff0c\u4f4e\u8d28\u91cf\uff0cJPEG\u538b\u7f29\u6b8b\u7559\uff0c\u4e11\u964b\u7684\uff0c\u6b8b\u7f3a\u7684\uff0c\u591a\u4f59\u7684\u624b\u6307\uff0c\u753b\u5f97\u4e0d\u597d\u7684\u624b\u90e8\uff0c\u753b\u5f97\u4e0d\u597d\u7684\u8138\u90e8\uff0c\u7578\u5f62\u7684\uff0c\u6bc1\u5bb9\u7684\uff0c\u5f62\u6001\u7578\u5f62\u7684\u80a2\u4f53\uff0c\u624b\u6307\u878d\u5408\uff0c\u9759\u6b62\u4e0d\u52a8\u7684\u753b\u9762\uff0c\u6742\u4e71\u7684\u80cc\u666f\uff0c\u4e09\u6761\u817f\uff0c\u80cc\u666f\u4eba\u5f88\u591a\uff0c\u5012\u7740\u8d70", "state_dtype": "bfloat16" }, "trajectories": 1000 }, "checkpoint_blocks": true, "cube_shape": [ 4, 4, 8 ], "endpoint_formula": "mean((start_latent - start_sigma * student_velocity - target_latent)^2)", "endpoint_loss": true, "energy_floor": 1e-06, "forward_path": "teacher_driven_local_reconstruction_plus_separate_full_student_velocity_forward_both_under_student_sparse_block_selection", "gradient_scope": "local_reconstruction_plus_end_to_end_ode_within_student_quantized_path", "input_source": "cached_ode_start_latent_and_corresponding_timestep_sigma_without_renoising", "loss": "mse", "mlp_weight": 0.25, "objective": "attention_mlp_reconstruction_and_ode_endpoint", "ode_weight": 0.5, "protocol_version": 2, "reconstruction_normalization": "none", "reduction": "attention_and_mlp_each_layer_mean_plus_ode_mean_mse_weighted_sum", "reference": { "attention_backend": "TORCH_SDPA", "attention_kind": "dense", "frozen": true, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "linear_quantization": null, "weights_sha256": "aac925e262c5c2266456830496021a66598b03253c8e05b6fa048b1744e0f7a1" }, "student_initialization": { "checkpoint_step": 500, "compression_gate_initialization": "zero", "init_from": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0316-recon-mixed-dense-g5-all12-b16-1000-260921-080538/export-step500", "initial_weight_equality_with_reference": false, "kind": "dense_qat_weight_initialization" }, "student_steps": 3, "target": [ "self_attention_post_output_projection_pre_gate_residual", "ffn_pre_gate_residual", "cached_teacher_final_clean_latent" ], "trainable_modules": [ "to_q", "to_k", "to_v", "to_out", "norm_q", "norm_k", "attn2", "ffn", "to_gate_compress" ], "trainable_scope": "quantized_path", "training_schedule": { "flow_shift": 8.0, "indices": [ 0, 16, 32 ], "sigmas": [ 1.0, 0.9411764144897461, 0.800000011920929 ], "student_steps": 3, "teacher_steps": 48, "timesteps": [ 1000.0, 941.1763916015625, 800.0 ], "version": 1 } }, "training_kernel_family": "nanovideo_triton_sparse_bf16_mm" } }, "parent_quantization_aware": true }, "distillation": { "checkpoint_roles": [ "student", "critic" ], "frozen_roles": [ "teacher" ], "method": { "_target_": "fastvideo.train.methods.distribution_matching.dmd2.DMD2Method", "dmd_denoising_steps": [ 1000.0, 941.1763916015625, 800.0 ], "fake_score_betas": [ 0.0, 0.999 ], "fake_score_learning_rate": 4e-07, "fake_score_lr_scheduler": "constant", "generator_update_interval": 5, "max_timestep_ratio": 0.98, "min_timestep_ratio": 0.02, "real_score_guidance_scale": 3.5, "rollout_mode": "simulate", "shifted_ode_schedule": { "flow_shift": 8.0, "indices": [ 0, 16, 32 ], "sigmas": [ 1.0, 0.9411764144897461, 0.800000011920929 ], "student_steps": 3, "teacher_steps": 48, "timesteps": [ 1000.0, 941.1763916015625, 800.0 ], "version": 1 }, "warp_denoising_step": false }, "roles": { "critic": { "_target_": "fastvideo.train.models.wan.WanModel", "attention_backend": "TORCH_SDPA", "disable_custom_init_weights": true, "enable_gradient_checkpointing_type": "full", "flow_shift": 8.0, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "trainable": true }, "student": { "_target_": "fastvideo.train.models.wan.WanModel", "attention_backend": "VSA_QAT_TRAIN_C128", "enable_gradient_checkpointing_type": "full", "flow_shift": 8.0, "init_from": "/home/hal-ywu/git/vsqa/logs/fastvideo-v0352-weighted-reconode-c128-k3-b16-1000-260922-104437/export-step500", "trainable": true }, "teacher": { "_target_": "fastvideo.train.models.wan.WanModel", "attention_backend": "TORCH_SDPA", "disable_custom_init_weights": true, "flow_shift": 8.0, "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "trainable": false } } }, "kernel_source_commit": "e8b4630e0ee15be13d1e0f6823882b776edbe62e", "kernel_source_files": { "block_sparse_attn.py": "da77938f5e68822e57b7571507064199c62440d580efa9e3b14f9b5d1bc72199", "qat/nvfp4_utils.py": "70b613fca22817ca764496095efc73127515574a53e4b9fcc70bd56a63b3ac3e", "qat/quant_utils.py": "e450efca2c375d48889c9d647f5760c9fe90330436dfc623a5684ebe6a48080d", "qat/sparse/bf16_mm.py": "1c8bbf7ac573ccfa3d28b41d001c922cb261c1c74bcc8e83743e0da092c82ff5", "triton_kernels/block_sparse_attn.py": "715d4d51ff26f3edf9764b6708769376068f228bedd1f7db26aee93375203d83", "triton_kernels/index.py": "ce8beccf1f66554917765976820834fdf03a45ad0d1852e6da0b16c811d7b19d" }, "linear_quantization": { "activation_global_scale": "dynamic, 448*6/max|X| recomputed every forward", "backward": "whole-linear STE on the saved unquantized X and master W in the gradient dtype", "format": "NVFP4 W4A4: E2M1 values, 16-element blocks, FP8 E4M3 block scales, FP32 global scale", "forward": "flashinfer nvfp4_quantize + mm_fp4 (cutlass) real FP4 GEMM, BF16 output", "kind": "nvfp4_qat_train", "master_weights": "FP32", "scope": "student transformer only; teacher and critic load with quant_config dropped", "target_prefix_substrings": [ "ffn.fc_in", "ffn.fc_out", "to_q", "to_k", "to_v", "to_out", "self_attention.out_layer", "cross_attention.out_layer", "feed_forward.mlp.fc_in", "feed_forward.mlp.fc_out" ], "weight_global_scale": "dynamic, 448*6/max|W| recomputed every forward from the master weight" }, "model_architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 12, "num_layers": 30, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "model_kind": "Wan bidirectional T2V", "physical_block_tokens": 64, "precision": { "coarse": "BF16", "compute": "BF16", "fine": "NVFP4 fake-quant with BF16 matmul", "master_weights": "FP32", "pv_fp8": false, "quant_p": true, "quant_v": true, "quantize_coarse": false, "quantize_fine": true, "two_level_p": false, "use_global_sf_p": false, "use_global_sf_qkv": false }, "selection": { "effective_topk": "max(1, min(ceil((1-sparsity)*num_blocks), num_blocks))", "mode": "top-p", "sparsity": 0.9 }, "training_kernel_family": "nanovideo_triton_sparse_bf16_mm" } }