{ "attention": { "backend": "VSA_QAT_TRAIN_C256", "compression_gates": "trained", "cube_shape": [ 4, 8, 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": "2026-09-23T23:20:37.109103+00:00", "id": "ckpt_7d8545a5ff1747d3acbb3390a5365e28", "state": "training", "step": 100 }, "checkpoint_state": "inference_export", "code_provenance": { "commit": "f93b269aff9e5eaa014b2582eab6ef79f061315a", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/mnt/lustre/vlm-miz055/vsqa-14b/code/fastvideo-f93b269-14b-dense5-260923" } }, "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": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0362-14b-c256-dmd1000-260922-123321/config/validation_16.json", "every_steps": 50, "guidance_scale": 1.0, "height": 768, "num_frames": 77, "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": 1280 } }, "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": 1, "max_timestep_ratio": 0.98, "min_timestep_ratio": 0.02, "real_score_guidance_scale": 6.0, "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-14B-Diffusers", "trainable": true }, "student": { "_target_": "fastvideo.train.models.wan.WanModel", "attention_backend": "VSA_QAT_TRAIN_C256", "enable_gradient_checkpointing_type": "full", "flow_shift": 8.0, "init_from": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0361-14b-c256-ode500-260922-123321/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-14B-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": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0362-14b-c256-dmd1000-260922-123321/checkpoints", "resume_from_checkpoint": "", "training_state_checkpointing_steps": 50 }, "data": { "data_path": "/mnt/lustre/vlm-miz055/vsqa-14b/data/vidprom-14b-conditioning-full-canonical-r38ec498/training_dataset", "dataloader_num_workers": 0, "num_frames": 77, "num_height": 768, "num_latent_t": 20, "num_width": 1280, "preprocessed_data_type": "text_only", "seed": 1000, "train_batch_size": 1, "training_cfg_rate": 0.0 }, "distributed": { "hsdp_replicate_dim": 1, "hsdp_shard_dim": 16, "num_gpus": 16, "sp_size": 4, "tp_size": 1 }, "dit_precision": "fp32", "loop": { "gradient_accumulation_steps": 4, "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-v0362-14b-c256-dmd1000-260922-123321", "trackers": [ "wandb" ] }, "vsa": { "cache_tile_buf": false, "sparsity": 0.9 } } }, "config_sha256": { "resolved": "5f2aeb738d8b4eaa6838087f6e7c81bb6ac81a18bd492d329f6b3603c3be24a6", "source": "b9910697697cb4f1e98e8646266531f2e3d630cdd9ebebc9e2f2a06d09d7d935" }, "metadata_schema": "fastvideo.nanovideo_vsa_training", "metadata_schema_version": 2, "model": { "architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "kind": "wan-t2v-14b" }, "provenance": { "code": { "commit": "f93b269aff9e5eaa014b2582eab6ef79f061315a", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/mnt/lustre/vlm-miz055/vsqa-14b/code/fastvideo-f93b269-14b-dense5-260923" } }, "configs": { "resolved": { "path": "resolved.yaml", "sha256": "5f2aeb738d8b4eaa6838087f6e7c81bb6ac81a18bd492d329f6b3603c3be24a6" }, "source": { "path": "source.yaml", "sha256": "b9910697697cb4f1e98e8646266531f2e3d630cdd9ebebc9e2f2a06d09d7d935" } }, "lineage": { "checkpoint_id": "ckpt_1dcccaae803c4fb6b1b69bebe85333f6", "checkpoint_step": 500, "compression_gate_initialization": "inherited", "init_from": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0361-14b-c256-ode500-260922-123321/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": 13824, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "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_C256", "parent_cube_shape": [ 4, 8, 8 ], "parent_metadata": { "attention": { "backend": "VSA_QAT_TRAIN_C256", "compression_gates": "trained", "cube_shape": [ 4, 8, 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": "2026-09-23T16:03:08.671087+00:00", "id": "ckpt_1dcccaae803c4fb6b1b69bebe85333f6", "state": "training", "step": 500 }, "checkpoint_state": "inference_export", "code_provenance": { "commit": "f93b269aff9e5eaa014b2582eab6ef79f061315a", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/mnt/lustre/vlm-miz055/vsqa-14b/code/fastvideo-f93b269-14b-dense5-260923" } }, "config": { "callbacks": { "grad_clip": { "_target_": "fastvideo.train.callbacks.grad_clip.GradNormClipCallback", "max_grad_norm": 10.0 }, "validation": { "_target_": "fastvideo.train.methods.ode_init.validation.ODEValidationCallback", "attn_qat_infer": false, "dataset_file": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0361-14b-c256-ode500-260922-123321/config/validation_16.json", "every_steps": 50, "guidance_scale": 1.0, "height": 768, "num_frames": 77, "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": 1280 } }, "method": { "_target_": "fastvideo.train.methods.ode_init.method.ODEInitMethod", "student_steps": 3 }, "models": { "student": { "_target_": "fastvideo.train.models.wan.WanModel", "attention_backend": "VSA_QAT_TRAIN_C256", "enable_gradient_checkpointing_type": "full", "flow_shift": 8.0, "init_from": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/export-step250", "trainable": true } }, "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": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0361-14b-c256-ode500-260922-123321/checkpoints", "resume_from_checkpoint": "", "training_state_checkpointing_steps": 50 }, "data": { "data_path": "/mnt/lustre/vlm-miz055/vsqa-14b/data/ode-init-vidprom-14b-250-48s12-77f-260922", "dataloader_num_workers": 0, "num_frames": 77, "num_height": 768, "num_latent_t": 20, "num_width": 1280, "seed": 1000, "train_batch_size": 1, "training_cfg_rate": 0.0 }, "distributed": { "hsdp_replicate_dim": 1, "hsdp_shard_dim": 16, "num_gpus": 16, "sp_size": 4, "tp_size": 1 }, "dit_precision": "fp32", "loop": { "gradient_accumulation_steps": 4, "max_train_steps": 500 }, "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-v0361-14b-c256-ode500-260922-123321", "trackers": [ "wandb" ] }, "vsa": { "cache_tile_buf": false, "sparsity": 0.9 } } }, "config_sha256": { "resolved": "b301b3d5727d4b19add1785711d70b3c6e068d4de8569329e864ee79bd570f61", "source": "adc72b6faf9be83741bef901f0aeb0b32f33a14ac55714e4360e03e8094be031" }, "metadata_schema": "fastvideo.nanovideo_vsa_training", "metadata_schema_version": 2, "model": { "architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "kind": "wan-t2v-14b" }, "provenance": { "code": { "commit": "f93b269aff9e5eaa014b2582eab6ef79f061315a", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/mnt/lustre/vlm-miz055/vsqa-14b/code/fastvideo-f93b269-14b-dense5-260923" } }, "configs": { "resolved": { "path": "resolved.yaml", "sha256": "b301b3d5727d4b19add1785711d70b3c6e068d4de8569329e864ee79bd570f61" }, "source": { "path": "source.yaml", "sha256": "adc72b6faf9be83741bef901f0aeb0b32f33a14ac55714e4360e03e8094be031" } }, "lineage": { "checkpoint_id": "ckpt_bbb601ff69f64efc8134f6624d59d074", "checkpoint_step": 250, "compression_gate_initialization": "zero", "init_from": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/export-step250", "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": 13824, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "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": "2026-09-23T10:50:41.063299+00:00", "id": "ckpt_bbb601ff69f64efc8134f6624d59d074", "state": "training", "step": 250 }, "checkpoint_state": "training_dcp", "code_provenance": { "commit": "f93b269aff9e5eaa014b2582eab6ef79f061315a", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/mnt/lustre/vlm-miz055/vsqa-14b/code/fastvideo-f93b269-14b-teacher-once-260923" } }, "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": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/config/validation_16.json", "dit_cpu_offload": false, "dit_layerwise_offload": false, "every_steps": 50, "guidance_scale": 1.0, "height": 768, "image_encoder_cpu_offload": false, "num_frames": 77, "offload_training_state": false, "pipeline_target": "fastvideo.pipelines.basic.wan.wan_shifted_dmd_pipeline.WanShiftedDMDPipeline", "run_at_start": false, "sampling_steps": [ 3 ], "text_encoder_cpu_offload": false, "unload_pipeline_after_validation": false, "use_validation_media_conditioning": false, "vae_cpu_offload": false, "width": 1280 } }, "method": { "_target_": "fastvideo.train.methods.layer_reconstruction.method.LayerReconstructionMethod", "checkpoint_blocks": true, "energy_floor": 1e-06, "loss": "mse", "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-14B-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-14B-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": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/checkpoints", "resume_from_checkpoint": "", "training_state_checkpointing_steps": 50 }, "data": { "data_path": "/mnt/lustre/vlm-miz055/vsqa-14b/data/ode-init-vidprom-14b-250-48s12-77f-260922", "dataloader_num_workers": 0, "num_frames": 77, "num_height": 768, "num_latent_t": 20, "num_width": 1280, "seed": 1000, "train_batch_size": 1, "training_cfg_rate": 0.0 }, "distributed": { "hsdp_replicate_dim": 1, "hsdp_shard_dim": 16, "num_gpus": 16, "sp_size": 4, "tp_size": 1 }, "dit_precision": "fp32", "loop": { "gradient_accumulation_steps": 4, "max_train_steps": 250 }, "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": 0.0 }, "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321", "trackers": [ "wandb" ] }, "vsa": { "cache_tile_buf": false, "sparsity": 0.0 } } }, "config_sha256": { "resolved": "ddd439e6333e8932194555ef24c95a695c67b465c3429f63e8683212c0fe365f", "source": "2f6a9099230711a0a5594e54fce56bafa1004596bbf4fb61581f667806481435" }, "metadata_schema": "fastvideo.dense_nvfp4_training", "metadata_schema_version": 2, "model": { "architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "kind": "wan-t2v-14b" }, "provenance": { "code": { "commit": "f93b269aff9e5eaa014b2582eab6ef79f061315a", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/mnt/lustre/vlm-miz055/vsqa-14b/code/fastvideo-f93b269-14b-teacher-once-260923" } }, "configs": { "resolved": { "path": "resolved.yaml", "sha256": "ddd439e6333e8932194555ef24c95a695c67b465c3429f63e8683212c0fe365f" }, "source": { "path": "source.yaml", "sha256": "2f6a9099230711a0a5594e54fce56bafa1004596bbf4fb61581f667806481435" } }, "lineage": { "compression_gate_initialization": null, "init_from": "Wan-AI/Wan2.1-T2V-14B-Diffusers", "kind": "base_model_initialization", "optimizer_state": "fresh" } }, "run_identity": { "output_dir": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/checkpoints", "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321", "trackers": [ "wandb" ] }, "wandb": { "id": "ys3z3hxi", "url": null } }, "step": 250, "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-14B-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": 40, "num_layers": 40, "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": 77, "num_height": 768, "num_latent_t": 20, "num_width": 1280 }, "pair_sampling": "all_twelve_teacher_anchors_to_clean", "pairs": 3000, "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": "vidprom", "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": 13824, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "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": "2992fe554957b20824172ebcb91070286fa4a35d94bd9a1c806cee136539719c", "transformer/diffusion_pytorch_model-00001-of-00012.safetensors": "63eb07f03808d8ec9b6bfc1b4f4968d7d1253f2569fec18fae114b4530554181", "transformer/diffusion_pytorch_model-00002-of-00012.safetensors": "8d24b40a5ea6bb21cdc0115128241ffeb44184b2419f4b1585af1573ed5c7839", "transformer/diffusion_pytorch_model-00003-of-00012.safetensors": "7f550f93b3c36c13015dd551109ff63528251864c9513acc2e0a502807b8a91c", "transformer/diffusion_pytorch_model-00004-of-00012.safetensors": "6a701d601e2cbe3da3515b24d59f56deb5ca2c973c778ff7911f579ec5b8ec11", "transformer/diffusion_pytorch_model-00005-of-00012.safetensors": "7eb65433e529b975f84897f6572af1ae148d82e7fd0ee7c5f4191158736b104f", "transformer/diffusion_pytorch_model-00006-of-00012.safetensors": "aae199dea0aeb96347e9af9e3cb90679c1c8c2f42c76ee3cead2ea240c909d07", "transformer/diffusion_pytorch_model-00007-of-00012.safetensors": "53c24e663dece6ff0737b03e5cc670118998231877ba7bad29824ff0500589d4", "transformer/diffusion_pytorch_model-00008-of-00012.safetensors": "333eaca9dce56bf3083bbab4630aa081b7c403a27ba94894a7756d631529fd9a", "transformer/diffusion_pytorch_model-00009-of-00012.safetensors": "8af811e7ed2ab64a3f6d07827d1b975a9804b78d63c2b3989024bc1b67ac6e5d", "transformer/diffusion_pytorch_model-00010-of-00012.safetensors": "e8b631e38f0ad21a63c631069658d4be4b55e0cb15e8f5848567482b5862ba63", "transformer/diffusion_pytorch_model-00011-of-00012.safetensors": "21032fe150df6f3ac26780fb8ca38d05c4a9585a1b774ceadb5d42bc96ff1b27", "transformer/diffusion_pytorch_model-00012-of-00012.safetensors": "7b36878bf76c01410c3be92e5239a10106dfdc932567553228879979e0ac903d", "vae/config.json": "f0c1cc1d7decb5badc384f54691746a27a9aeff49f7ebca974e583389342d527" }, "guidance_scale": 6.0, "model_kind": "wan-t2v-14b", "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": 250 }, "checkpoint_blocks": true, "endpoint_loss": false, "energy_floor": 1e-06, "forward_path": "fixed_teacher_blocks_with_identical_frozen_student_embedding_and_conditioning", "gradient_scope": "block_local_student_quantized_path", "input_source": "cached_ode_start_latent_and_corresponding_timestep_sigma_without_renoising", "loss": "mse", "normalized_mse_denominator": "detached_per_sample_reference_block_update_mean_square_clamped_to_energy_floor", "objective": "fixed_teacher_layer_reconstruction", "protocol_version": 1, "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-14B-Diffusers", "linear_quantization": null, "weights_sha256": "33fd981340806b2cccd01dc431cefc7334022c2cc6c67980c4a06c2a82fe88bb" }, "target": "transformer_block_output_hidden_state", "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": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0361-14b-c256-ode500-260922-123321/checkpoints", "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0361-14b-c256-ode500-260922-123321", "trackers": [ "wandb" ] }, "wandb": { "id": "d8mbh20r", "url": null } }, "step": 500, "training_contract": { "attention_backend": "VSA_QAT_TRAIN_C256", "attention_kind": "VSA", "cube_shape": [ 4, 8, 8 ], "direct_lineage": { "checkpoint_id": "ckpt_bbb601ff69f64efc8134f6624d59d074", "checkpoint_step": 250, "compression_gate_initialization": "zero", "init_from": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/export-step250", "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": 13824, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "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": "2026-09-23T10:50:41.063299+00:00", "id": "ckpt_bbb601ff69f64efc8134f6624d59d074", "state": "training", "step": 250 }, "checkpoint_state": "training_dcp", "code_provenance": { "commit": "f93b269aff9e5eaa014b2582eab6ef79f061315a", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/mnt/lustre/vlm-miz055/vsqa-14b/code/fastvideo-f93b269-14b-teacher-once-260923" } }, "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": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/config/validation_16.json", "dit_cpu_offload": false, "dit_layerwise_offload": false, "every_steps": 50, "guidance_scale": 1.0, "height": 768, "image_encoder_cpu_offload": false, "num_frames": 77, "offload_training_state": false, "pipeline_target": "fastvideo.pipelines.basic.wan.wan_shifted_dmd_pipeline.WanShiftedDMDPipeline", "run_at_start": false, "sampling_steps": [ 3 ], "text_encoder_cpu_offload": false, "unload_pipeline_after_validation": false, "use_validation_media_conditioning": false, "vae_cpu_offload": false, "width": 1280 } }, "method": { "_target_": "fastvideo.train.methods.layer_reconstruction.method.LayerReconstructionMethod", "checkpoint_blocks": true, "energy_floor": 1e-06, "loss": "mse", "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-14B-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-14B-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": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/checkpoints", "resume_from_checkpoint": "", "training_state_checkpointing_steps": 50 }, "data": { "data_path": "/mnt/lustre/vlm-miz055/vsqa-14b/data/ode-init-vidprom-14b-250-48s12-77f-260922", "dataloader_num_workers": 0, "num_frames": 77, "num_height": 768, "num_latent_t": 20, "num_width": 1280, "seed": 1000, "train_batch_size": 1, "training_cfg_rate": 0.0 }, "distributed": { "hsdp_replicate_dim": 1, "hsdp_shard_dim": 16, "num_gpus": 16, "sp_size": 4, "tp_size": 1 }, "dit_precision": "fp32", "loop": { "gradient_accumulation_steps": 4, "max_train_steps": 250 }, "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": 0.0 }, "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321", "trackers": [ "wandb" ] }, "vsa": { "cache_tile_buf": false, "sparsity": 0.0 } } }, "config_sha256": { "resolved": "ddd439e6333e8932194555ef24c95a695c67b465c3429f63e8683212c0fe365f", "source": "2f6a9099230711a0a5594e54fce56bafa1004596bbf4fb61581f667806481435" }, "metadata_schema": "fastvideo.dense_nvfp4_training", "metadata_schema_version": 2, "model": { "architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "kind": "wan-t2v-14b" }, "provenance": { "code": { "commit": "f93b269aff9e5eaa014b2582eab6ef79f061315a", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/mnt/lustre/vlm-miz055/vsqa-14b/code/fastvideo-f93b269-14b-teacher-once-260923" } }, "configs": { "resolved": { "path": "resolved.yaml", "sha256": "ddd439e6333e8932194555ef24c95a695c67b465c3429f63e8683212c0fe365f" }, "source": { "path": "source.yaml", "sha256": "2f6a9099230711a0a5594e54fce56bafa1004596bbf4fb61581f667806481435" } }, "lineage": { "compression_gate_initialization": null, "init_from": "Wan-AI/Wan2.1-T2V-14B-Diffusers", "kind": "base_model_initialization", "optimizer_state": "fresh" } }, "run_identity": { "output_dir": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/checkpoints", "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321", "trackers": [ "wandb" ] }, "wandb": { "id": "ys3z3hxi", "url": null } }, "step": 250, "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-14B-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": 40, "num_layers": 40, "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": 77, "num_height": 768, "num_latent_t": 20, "num_width": 1280 }, "pair_sampling": "all_twelve_teacher_anchors_to_clean", "pairs": 3000, "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": "vidprom", "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": 13824, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "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": "2992fe554957b20824172ebcb91070286fa4a35d94bd9a1c806cee136539719c", "transformer/diffusion_pytorch_model-00001-of-00012.safetensors": "63eb07f03808d8ec9b6bfc1b4f4968d7d1253f2569fec18fae114b4530554181", "transformer/diffusion_pytorch_model-00002-of-00012.safetensors": "8d24b40a5ea6bb21cdc0115128241ffeb44184b2419f4b1585af1573ed5c7839", "transformer/diffusion_pytorch_model-00003-of-00012.safetensors": "7f550f93b3c36c13015dd551109ff63528251864c9513acc2e0a502807b8a91c", "transformer/diffusion_pytorch_model-00004-of-00012.safetensors": "6a701d601e2cbe3da3515b24d59f56deb5ca2c973c778ff7911f579ec5b8ec11", "transformer/diffusion_pytorch_model-00005-of-00012.safetensors": "7eb65433e529b975f84897f6572af1ae148d82e7fd0ee7c5f4191158736b104f", "transformer/diffusion_pytorch_model-00006-of-00012.safetensors": "aae199dea0aeb96347e9af9e3cb90679c1c8c2f42c76ee3cead2ea240c909d07", "transformer/diffusion_pytorch_model-00007-of-00012.safetensors": "53c24e663dece6ff0737b03e5cc670118998231877ba7bad29824ff0500589d4", "transformer/diffusion_pytorch_model-00008-of-00012.safetensors": "333eaca9dce56bf3083bbab4630aa081b7c403a27ba94894a7756d631529fd9a", "transformer/diffusion_pytorch_model-00009-of-00012.safetensors": "8af811e7ed2ab64a3f6d07827d1b975a9804b78d63c2b3989024bc1b67ac6e5d", "transformer/diffusion_pytorch_model-00010-of-00012.safetensors": "e8b631e38f0ad21a63c631069658d4be4b55e0cb15e8f5848567482b5862ba63", "transformer/diffusion_pytorch_model-00011-of-00012.safetensors": "21032fe150df6f3ac26780fb8ca38d05c4a9585a1b774ceadb5d42bc96ff1b27", "transformer/diffusion_pytorch_model-00012-of-00012.safetensors": "7b36878bf76c01410c3be92e5239a10106dfdc932567553228879979e0ac903d", "vae/config.json": "f0c1cc1d7decb5badc384f54691746a27a9aeff49f7ebca974e583389342d527" }, "guidance_scale": 6.0, "model_kind": "wan-t2v-14b", "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": 250 }, "checkpoint_blocks": true, "endpoint_loss": false, "energy_floor": 1e-06, "forward_path": "fixed_teacher_blocks_with_identical_frozen_student_embedding_and_conditioning", "gradient_scope": "block_local_student_quantized_path", "input_source": "cached_ode_start_latent_and_corresponding_timestep_sigma_without_renoising", "loss": "mse", "normalized_mse_denominator": "detached_per_sample_reference_block_update_mean_square_clamped_to_energy_floor", "objective": "fixed_teacher_layer_reconstruction", "protocol_version": 1, "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-14B-Diffusers", "linear_quantization": null, "weights_sha256": "33fd981340806b2cccd01dc431cefc7334022c2cc6c67980c4a06c2a82fe88bb" }, "target": "transformer_block_output_hidden_state", "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; 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": 40, "num_layers": 40, "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": { "base_seed": 1000, "conditioning": "teacher_cfg", "end_timestep_conditioning": false, "geometry": { "num_frames": 77, "num_height": 768, "num_latent_t": 20, "num_width": 1280 }, "objective": "teacher_trajectory_endpoint_mse", "pair_sampling": "all_twelve_teacher_anchors_to_clean", "pairs": 3000, "pairs_per_trajectory": 12, "parameterization": "flow_velocity_to_clean_latent", "reduction": "float32_mean", "rng": "sha256_sample_stream_v1_fixed_endpoint_pairs", "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": "vidprom", "target_source": "offline_pure_noise_teacher_trajectory", "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": 13824, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "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": "2992fe554957b20824172ebcb91070286fa4a35d94bd9a1c806cee136539719c", "transformer/diffusion_pytorch_model-00001-of-00012.safetensors": "63eb07f03808d8ec9b6bfc1b4f4968d7d1253f2569fec18fae114b4530554181", "transformer/diffusion_pytorch_model-00002-of-00012.safetensors": "8d24b40a5ea6bb21cdc0115128241ffeb44184b2419f4b1585af1573ed5c7839", "transformer/diffusion_pytorch_model-00003-of-00012.safetensors": "7f550f93b3c36c13015dd551109ff63528251864c9513acc2e0a502807b8a91c", "transformer/diffusion_pytorch_model-00004-of-00012.safetensors": "6a701d601e2cbe3da3515b24d59f56deb5ca2c973c778ff7911f579ec5b8ec11", "transformer/diffusion_pytorch_model-00005-of-00012.safetensors": "7eb65433e529b975f84897f6572af1ae148d82e7fd0ee7c5f4191158736b104f", "transformer/diffusion_pytorch_model-00006-of-00012.safetensors": "aae199dea0aeb96347e9af9e3cb90679c1c8c2f42c76ee3cead2ea240c909d07", "transformer/diffusion_pytorch_model-00007-of-00012.safetensors": "53c24e663dece6ff0737b03e5cc670118998231877ba7bad29824ff0500589d4", "transformer/diffusion_pytorch_model-00008-of-00012.safetensors": "333eaca9dce56bf3083bbab4630aa081b7c403a27ba94894a7756d631529fd9a", "transformer/diffusion_pytorch_model-00009-of-00012.safetensors": "8af811e7ed2ab64a3f6d07827d1b975a9804b78d63c2b3989024bc1b67ac6e5d", "transformer/diffusion_pytorch_model-00010-of-00012.safetensors": "e8b631e38f0ad21a63c631069658d4be4b55e0cb15e8f5848567482b5862ba63", "transformer/diffusion_pytorch_model-00011-of-00012.safetensors": "21032fe150df6f3ac26780fb8ca38d05c4a9585a1b774ceadb5d42bc96ff1b27", "transformer/diffusion_pytorch_model-00012-of-00012.safetensors": "7b36878bf76c01410c3be92e5239a10106dfdc932567553228879979e0ac903d", "vae/config.json": "f0c1cc1d7decb5badc384f54691746a27a9aeff49f7ebca974e583389342d527" }, "guidance_scale": 6.0, "model_kind": "wan-t2v-14b", "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" }, "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 }, "trajectories": 250 }, "training_kernel_family": "nanovideo_triton_sparse_bf16_mm" } }, "parent_quantization_aware": true } }, "run_identity": { "output_dir": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0362-14b-c256-dmd1000-260922-123321/checkpoints", "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0362-14b-c256-dmd1000-260922-123321", "trackers": [ "wandb" ] }, "wandb": { "id": "dmlnbl5x", "url": null } }, "step": 100, "training_contract": { "attention_backend": "VSA_QAT_TRAIN_C256", "attention_kind": "VSA", "cube_shape": [ 4, 8, 8 ], "direct_lineage": { "checkpoint_id": "ckpt_1dcccaae803c4fb6b1b69bebe85333f6", "checkpoint_step": 500, "compression_gate_initialization": "inherited", "init_from": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0361-14b-c256-ode500-260922-123321/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": 13824, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "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_C256", "parent_cube_shape": [ 4, 8, 8 ], "parent_metadata": { "attention": { "backend": "VSA_QAT_TRAIN_C256", "compression_gates": "trained", "cube_shape": [ 4, 8, 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": "2026-09-23T16:03:08.671087+00:00", "id": "ckpt_1dcccaae803c4fb6b1b69bebe85333f6", "state": "training", "step": 500 }, "checkpoint_state": "inference_export", "code_provenance": { "commit": "f93b269aff9e5eaa014b2582eab6ef79f061315a", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/mnt/lustre/vlm-miz055/vsqa-14b/code/fastvideo-f93b269-14b-dense5-260923" } }, "config": { "callbacks": { "grad_clip": { "_target_": "fastvideo.train.callbacks.grad_clip.GradNormClipCallback", "max_grad_norm": 10.0 }, "validation": { "_target_": "fastvideo.train.methods.ode_init.validation.ODEValidationCallback", "attn_qat_infer": false, "dataset_file": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0361-14b-c256-ode500-260922-123321/config/validation_16.json", "every_steps": 50, "guidance_scale": 1.0, "height": 768, "num_frames": 77, "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": 1280 } }, "method": { "_target_": "fastvideo.train.methods.ode_init.method.ODEInitMethod", "student_steps": 3 }, "models": { "student": { "_target_": "fastvideo.train.models.wan.WanModel", "attention_backend": "VSA_QAT_TRAIN_C256", "enable_gradient_checkpointing_type": "full", "flow_shift": 8.0, "init_from": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/export-step250", "trainable": true } }, "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": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0361-14b-c256-ode500-260922-123321/checkpoints", "resume_from_checkpoint": "", "training_state_checkpointing_steps": 50 }, "data": { "data_path": "/mnt/lustre/vlm-miz055/vsqa-14b/data/ode-init-vidprom-14b-250-48s12-77f-260922", "dataloader_num_workers": 0, "num_frames": 77, "num_height": 768, "num_latent_t": 20, "num_width": 1280, "seed": 1000, "train_batch_size": 1, "training_cfg_rate": 0.0 }, "distributed": { "hsdp_replicate_dim": 1, "hsdp_shard_dim": 16, "num_gpus": 16, "sp_size": 4, "tp_size": 1 }, "dit_precision": "fp32", "loop": { "gradient_accumulation_steps": 4, "max_train_steps": 500 }, "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-v0361-14b-c256-ode500-260922-123321", "trackers": [ "wandb" ] }, "vsa": { "cache_tile_buf": false, "sparsity": 0.9 } } }, "config_sha256": { "resolved": "b301b3d5727d4b19add1785711d70b3c6e068d4de8569329e864ee79bd570f61", "source": "adc72b6faf9be83741bef901f0aeb0b32f33a14ac55714e4360e03e8094be031" }, "metadata_schema": "fastvideo.nanovideo_vsa_training", "metadata_schema_version": 2, "model": { "architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "kind": "wan-t2v-14b" }, "provenance": { "code": { "commit": "f93b269aff9e5eaa014b2582eab6ef79f061315a", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/mnt/lustre/vlm-miz055/vsqa-14b/code/fastvideo-f93b269-14b-dense5-260923" } }, "configs": { "resolved": { "path": "resolved.yaml", "sha256": "b301b3d5727d4b19add1785711d70b3c6e068d4de8569329e864ee79bd570f61" }, "source": { "path": "source.yaml", "sha256": "adc72b6faf9be83741bef901f0aeb0b32f33a14ac55714e4360e03e8094be031" } }, "lineage": { "checkpoint_id": "ckpt_bbb601ff69f64efc8134f6624d59d074", "checkpoint_step": 250, "compression_gate_initialization": "zero", "init_from": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/export-step250", "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": 13824, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "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": "2026-09-23T10:50:41.063299+00:00", "id": "ckpt_bbb601ff69f64efc8134f6624d59d074", "state": "training", "step": 250 }, "checkpoint_state": "training_dcp", "code_provenance": { "commit": "f93b269aff9e5eaa014b2582eab6ef79f061315a", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/mnt/lustre/vlm-miz055/vsqa-14b/code/fastvideo-f93b269-14b-teacher-once-260923" } }, "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": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/config/validation_16.json", "dit_cpu_offload": false, "dit_layerwise_offload": false, "every_steps": 50, "guidance_scale": 1.0, "height": 768, "image_encoder_cpu_offload": false, "num_frames": 77, "offload_training_state": false, "pipeline_target": "fastvideo.pipelines.basic.wan.wan_shifted_dmd_pipeline.WanShiftedDMDPipeline", "run_at_start": false, "sampling_steps": [ 3 ], "text_encoder_cpu_offload": false, "unload_pipeline_after_validation": false, "use_validation_media_conditioning": false, "vae_cpu_offload": false, "width": 1280 } }, "method": { "_target_": "fastvideo.train.methods.layer_reconstruction.method.LayerReconstructionMethod", "checkpoint_blocks": true, "energy_floor": 1e-06, "loss": "mse", "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-14B-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-14B-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": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/checkpoints", "resume_from_checkpoint": "", "training_state_checkpointing_steps": 50 }, "data": { "data_path": "/mnt/lustre/vlm-miz055/vsqa-14b/data/ode-init-vidprom-14b-250-48s12-77f-260922", "dataloader_num_workers": 0, "num_frames": 77, "num_height": 768, "num_latent_t": 20, "num_width": 1280, "seed": 1000, "train_batch_size": 1, "training_cfg_rate": 0.0 }, "distributed": { "hsdp_replicate_dim": 1, "hsdp_shard_dim": 16, "num_gpus": 16, "sp_size": 4, "tp_size": 1 }, "dit_precision": "fp32", "loop": { "gradient_accumulation_steps": 4, "max_train_steps": 250 }, "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": 0.0 }, "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321", "trackers": [ "wandb" ] }, "vsa": { "cache_tile_buf": false, "sparsity": 0.0 } } }, "config_sha256": { "resolved": "ddd439e6333e8932194555ef24c95a695c67b465c3429f63e8683212c0fe365f", "source": "2f6a9099230711a0a5594e54fce56bafa1004596bbf4fb61581f667806481435" }, "metadata_schema": "fastvideo.dense_nvfp4_training", "metadata_schema_version": 2, "model": { "architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "kind": "wan-t2v-14b" }, "provenance": { "code": { "commit": "f93b269aff9e5eaa014b2582eab6ef79f061315a", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/mnt/lustre/vlm-miz055/vsqa-14b/code/fastvideo-f93b269-14b-teacher-once-260923" } }, "configs": { "resolved": { "path": "resolved.yaml", "sha256": "ddd439e6333e8932194555ef24c95a695c67b465c3429f63e8683212c0fe365f" }, "source": { "path": "source.yaml", "sha256": "2f6a9099230711a0a5594e54fce56bafa1004596bbf4fb61581f667806481435" } }, "lineage": { "compression_gate_initialization": null, "init_from": "Wan-AI/Wan2.1-T2V-14B-Diffusers", "kind": "base_model_initialization", "optimizer_state": "fresh" } }, "run_identity": { "output_dir": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/checkpoints", "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321", "trackers": [ "wandb" ] }, "wandb": { "id": "ys3z3hxi", "url": null } }, "step": 250, "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-14B-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": 40, "num_layers": 40, "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": 77, "num_height": 768, "num_latent_t": 20, "num_width": 1280 }, "pair_sampling": "all_twelve_teacher_anchors_to_clean", "pairs": 3000, "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": "vidprom", "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": 13824, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "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": "2992fe554957b20824172ebcb91070286fa4a35d94bd9a1c806cee136539719c", "transformer/diffusion_pytorch_model-00001-of-00012.safetensors": "63eb07f03808d8ec9b6bfc1b4f4968d7d1253f2569fec18fae114b4530554181", "transformer/diffusion_pytorch_model-00002-of-00012.safetensors": "8d24b40a5ea6bb21cdc0115128241ffeb44184b2419f4b1585af1573ed5c7839", "transformer/diffusion_pytorch_model-00003-of-00012.safetensors": "7f550f93b3c36c13015dd551109ff63528251864c9513acc2e0a502807b8a91c", "transformer/diffusion_pytorch_model-00004-of-00012.safetensors": "6a701d601e2cbe3da3515b24d59f56deb5ca2c973c778ff7911f579ec5b8ec11", "transformer/diffusion_pytorch_model-00005-of-00012.safetensors": "7eb65433e529b975f84897f6572af1ae148d82e7fd0ee7c5f4191158736b104f", "transformer/diffusion_pytorch_model-00006-of-00012.safetensors": "aae199dea0aeb96347e9af9e3cb90679c1c8c2f42c76ee3cead2ea240c909d07", "transformer/diffusion_pytorch_model-00007-of-00012.safetensors": "53c24e663dece6ff0737b03e5cc670118998231877ba7bad29824ff0500589d4", "transformer/diffusion_pytorch_model-00008-of-00012.safetensors": "333eaca9dce56bf3083bbab4630aa081b7c403a27ba94894a7756d631529fd9a", "transformer/diffusion_pytorch_model-00009-of-00012.safetensors": "8af811e7ed2ab64a3f6d07827d1b975a9804b78d63c2b3989024bc1b67ac6e5d", "transformer/diffusion_pytorch_model-00010-of-00012.safetensors": "e8b631e38f0ad21a63c631069658d4be4b55e0cb15e8f5848567482b5862ba63", "transformer/diffusion_pytorch_model-00011-of-00012.safetensors": "21032fe150df6f3ac26780fb8ca38d05c4a9585a1b774ceadb5d42bc96ff1b27", "transformer/diffusion_pytorch_model-00012-of-00012.safetensors": "7b36878bf76c01410c3be92e5239a10106dfdc932567553228879979e0ac903d", "vae/config.json": "f0c1cc1d7decb5badc384f54691746a27a9aeff49f7ebca974e583389342d527" }, "guidance_scale": 6.0, "model_kind": "wan-t2v-14b", "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": 250 }, "checkpoint_blocks": true, "endpoint_loss": false, "energy_floor": 1e-06, "forward_path": "fixed_teacher_blocks_with_identical_frozen_student_embedding_and_conditioning", "gradient_scope": "block_local_student_quantized_path", "input_source": "cached_ode_start_latent_and_corresponding_timestep_sigma_without_renoising", "loss": "mse", "normalized_mse_denominator": "detached_per_sample_reference_block_update_mean_square_clamped_to_energy_floor", "objective": "fixed_teacher_layer_reconstruction", "protocol_version": 1, "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-14B-Diffusers", "linear_quantization": null, "weights_sha256": "33fd981340806b2cccd01dc431cefc7334022c2cc6c67980c4a06c2a82fe88bb" }, "target": "transformer_block_output_hidden_state", "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": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0361-14b-c256-ode500-260922-123321/checkpoints", "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0361-14b-c256-ode500-260922-123321", "trackers": [ "wandb" ] }, "wandb": { "id": "d8mbh20r", "url": null } }, "step": 500, "training_contract": { "attention_backend": "VSA_QAT_TRAIN_C256", "attention_kind": "VSA", "cube_shape": [ 4, 8, 8 ], "direct_lineage": { "checkpoint_id": "ckpt_bbb601ff69f64efc8134f6624d59d074", "checkpoint_step": 250, "compression_gate_initialization": "zero", "init_from": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/export-step250", "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": 13824, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "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": "2026-09-23T10:50:41.063299+00:00", "id": "ckpt_bbb601ff69f64efc8134f6624d59d074", "state": "training", "step": 250 }, "checkpoint_state": "training_dcp", "code_provenance": { "commit": "f93b269aff9e5eaa014b2582eab6ef79f061315a", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/mnt/lustre/vlm-miz055/vsqa-14b/code/fastvideo-f93b269-14b-teacher-once-260923" } }, "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": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/config/validation_16.json", "dit_cpu_offload": false, "dit_layerwise_offload": false, "every_steps": 50, "guidance_scale": 1.0, "height": 768, "image_encoder_cpu_offload": false, "num_frames": 77, "offload_training_state": false, "pipeline_target": "fastvideo.pipelines.basic.wan.wan_shifted_dmd_pipeline.WanShiftedDMDPipeline", "run_at_start": false, "sampling_steps": [ 3 ], "text_encoder_cpu_offload": false, "unload_pipeline_after_validation": false, "use_validation_media_conditioning": false, "vae_cpu_offload": false, "width": 1280 } }, "method": { "_target_": "fastvideo.train.methods.layer_reconstruction.method.LayerReconstructionMethod", "checkpoint_blocks": true, "energy_floor": 1e-06, "loss": "mse", "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-14B-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-14B-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": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/checkpoints", "resume_from_checkpoint": "", "training_state_checkpointing_steps": 50 }, "data": { "data_path": "/mnt/lustre/vlm-miz055/vsqa-14b/data/ode-init-vidprom-14b-250-48s12-77f-260922", "dataloader_num_workers": 0, "num_frames": 77, "num_height": 768, "num_latent_t": 20, "num_width": 1280, "seed": 1000, "train_batch_size": 1, "training_cfg_rate": 0.0 }, "distributed": { "hsdp_replicate_dim": 1, "hsdp_shard_dim": 16, "num_gpus": 16, "sp_size": 4, "tp_size": 1 }, "dit_precision": "fp32", "loop": { "gradient_accumulation_steps": 4, "max_train_steps": 250 }, "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": 0.0 }, "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321", "trackers": [ "wandb" ] }, "vsa": { "cache_tile_buf": false, "sparsity": 0.0 } } }, "config_sha256": { "resolved": "ddd439e6333e8932194555ef24c95a695c67b465c3429f63e8683212c0fe365f", "source": "2f6a9099230711a0a5594e54fce56bafa1004596bbf4fb61581f667806481435" }, "metadata_schema": "fastvideo.dense_nvfp4_training", "metadata_schema_version": 2, "model": { "architecture": { "attention_head_dim": 128, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "out_channels": 16, "patch_size": [ 1, 2, 2 ] }, "kind": "wan-t2v-14b" }, "provenance": { "code": { "commit": "f93b269aff9e5eaa014b2582eab6ef79f061315a", "dirty": true, "policy": "path-and-commit-no-manifest-hash", "repo": "fastvideo-internal-dev", "source_archive": { "root": "/mnt/lustre/vlm-miz055/vsqa-14b/code/fastvideo-f93b269-14b-teacher-once-260923" } }, "configs": { "resolved": { "path": "resolved.yaml", "sha256": "ddd439e6333e8932194555ef24c95a695c67b465c3429f63e8683212c0fe365f" }, "source": { "path": "source.yaml", "sha256": "2f6a9099230711a0a5594e54fce56bafa1004596bbf4fb61581f667806481435" } }, "lineage": { "compression_gate_initialization": null, "init_from": "Wan-AI/Wan2.1-T2V-14B-Diffusers", "kind": "base_model_initialization", "optimizer_state": "fresh" } }, "run_identity": { "output_dir": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321/checkpoints", "tracker": { "entity": "alexzms-ucsd", "project_name": "fastvideo", "run_name": "fastvideo-v0360-14b-dense-nvfp4-recon250-260922-123321", "trackers": [ "wandb" ] }, "wandb": { "id": "ys3z3hxi", "url": null } }, "step": 250, "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-14B-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": 40, "num_layers": 40, "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": 77, "num_height": 768, "num_latent_t": 20, "num_width": 1280 }, "pair_sampling": "all_twelve_teacher_anchors_to_clean", "pairs": 3000, "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": "vidprom", "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": 13824, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "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": "2992fe554957b20824172ebcb91070286fa4a35d94bd9a1c806cee136539719c", "transformer/diffusion_pytorch_model-00001-of-00012.safetensors": "63eb07f03808d8ec9b6bfc1b4f4968d7d1253f2569fec18fae114b4530554181", "transformer/diffusion_pytorch_model-00002-of-00012.safetensors": "8d24b40a5ea6bb21cdc0115128241ffeb44184b2419f4b1585af1573ed5c7839", "transformer/diffusion_pytorch_model-00003-of-00012.safetensors": "7f550f93b3c36c13015dd551109ff63528251864c9513acc2e0a502807b8a91c", "transformer/diffusion_pytorch_model-00004-of-00012.safetensors": "6a701d601e2cbe3da3515b24d59f56deb5ca2c973c778ff7911f579ec5b8ec11", "transformer/diffusion_pytorch_model-00005-of-00012.safetensors": "7eb65433e529b975f84897f6572af1ae148d82e7fd0ee7c5f4191158736b104f", "transformer/diffusion_pytorch_model-00006-of-00012.safetensors": "aae199dea0aeb96347e9af9e3cb90679c1c8c2f42c76ee3cead2ea240c909d07", "transformer/diffusion_pytorch_model-00007-of-00012.safetensors": "53c24e663dece6ff0737b03e5cc670118998231877ba7bad29824ff0500589d4", "transformer/diffusion_pytorch_model-00008-of-00012.safetensors": "333eaca9dce56bf3083bbab4630aa081b7c403a27ba94894a7756d631529fd9a", "transformer/diffusion_pytorch_model-00009-of-00012.safetensors": "8af811e7ed2ab64a3f6d07827d1b975a9804b78d63c2b3989024bc1b67ac6e5d", "transformer/diffusion_pytorch_model-00010-of-00012.safetensors": "e8b631e38f0ad21a63c631069658d4be4b55e0cb15e8f5848567482b5862ba63", "transformer/diffusion_pytorch_model-00011-of-00012.safetensors": "21032fe150df6f3ac26780fb8ca38d05c4a9585a1b774ceadb5d42bc96ff1b27", "transformer/diffusion_pytorch_model-00012-of-00012.safetensors": "7b36878bf76c01410c3be92e5239a10106dfdc932567553228879979e0ac903d", "vae/config.json": "f0c1cc1d7decb5badc384f54691746a27a9aeff49f7ebca974e583389342d527" }, "guidance_scale": 6.0, "model_kind": "wan-t2v-14b", "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": 250 }, "checkpoint_blocks": true, "endpoint_loss": false, "energy_floor": 1e-06, "forward_path": "fixed_teacher_blocks_with_identical_frozen_student_embedding_and_conditioning", "gradient_scope": "block_local_student_quantized_path", "input_source": "cached_ode_start_latent_and_corresponding_timestep_sigma_without_renoising", "loss": "mse", "normalized_mse_denominator": "detached_per_sample_reference_block_update_mean_square_clamped_to_energy_floor", "objective": "fixed_teacher_layer_reconstruction", "protocol_version": 1, "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-14B-Diffusers", "linear_quantization": null, "weights_sha256": "33fd981340806b2cccd01dc431cefc7334022c2cc6c67980c4a06c2a82fe88bb" }, "target": "transformer_block_output_hidden_state", "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; 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": 40, "num_layers": 40, "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": { "base_seed": 1000, "conditioning": "teacher_cfg", "end_timestep_conditioning": false, "geometry": { "num_frames": 77, "num_height": 768, "num_latent_t": 20, "num_width": 1280 }, "objective": "teacher_trajectory_endpoint_mse", "pair_sampling": "all_twelve_teacher_anchors_to_clean", "pairs": 3000, "pairs_per_trajectory": 12, "parameterization": "flow_velocity_to_clean_latent", "reduction": "float32_mean", "rng": "sha256_sample_stream_v1_fixed_endpoint_pairs", "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": "vidprom", "target_source": "offline_pure_noise_teacher_trajectory", "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": 13824, "freq_dim": 256, "image_dim": null, "in_channels": 16, "num_attention_heads": 40, "num_layers": 40, "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": "2992fe554957b20824172ebcb91070286fa4a35d94bd9a1c806cee136539719c", "transformer/diffusion_pytorch_model-00001-of-00012.safetensors": "63eb07f03808d8ec9b6bfc1b4f4968d7d1253f2569fec18fae114b4530554181", "transformer/diffusion_pytorch_model-00002-of-00012.safetensors": "8d24b40a5ea6bb21cdc0115128241ffeb44184b2419f4b1585af1573ed5c7839", "transformer/diffusion_pytorch_model-00003-of-00012.safetensors": "7f550f93b3c36c13015dd551109ff63528251864c9513acc2e0a502807b8a91c", "transformer/diffusion_pytorch_model-00004-of-00012.safetensors": "6a701d601e2cbe3da3515b24d59f56deb5ca2c973c778ff7911f579ec5b8ec11", "transformer/diffusion_pytorch_model-00005-of-00012.safetensors": "7eb65433e529b975f84897f6572af1ae148d82e7fd0ee7c5f4191158736b104f", "transformer/diffusion_pytorch_model-00006-of-00012.safetensors": "aae199dea0aeb96347e9af9e3cb90679c1c8c2f42c76ee3cead2ea240c909d07", "transformer/diffusion_pytorch_model-00007-of-00012.safetensors": "53c24e663dece6ff0737b03e5cc670118998231877ba7bad29824ff0500589d4", "transformer/diffusion_pytorch_model-00008-of-00012.safetensors": "333eaca9dce56bf3083bbab4630aa081b7c403a27ba94894a7756d631529fd9a", "transformer/diffusion_pytorch_model-00009-of-00012.safetensors": "8af811e7ed2ab64a3f6d07827d1b975a9804b78d63c2b3989024bc1b67ac6e5d", "transformer/diffusion_pytorch_model-00010-of-00012.safetensors": "e8b631e38f0ad21a63c631069658d4be4b55e0cb15e8f5848567482b5862ba63", "transformer/diffusion_pytorch_model-00011-of-00012.safetensors": "21032fe150df6f3ac26780fb8ca38d05c4a9585a1b774ceadb5d42bc96ff1b27", "transformer/diffusion_pytorch_model-00012-of-00012.safetensors": "7b36878bf76c01410c3be92e5239a10106dfdc932567553228879979e0ac903d", "vae/config.json": "f0c1cc1d7decb5badc384f54691746a27a9aeff49f7ebca974e583389342d527" }, "guidance_scale": 6.0, "model_kind": "wan-t2v-14b", "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" }, "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 }, "trajectories": 250 }, "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": 1, "max_timestep_ratio": 0.98, "min_timestep_ratio": 0.02, "real_score_guidance_scale": 6.0, "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-14B-Diffusers", "trainable": true }, "student": { "_target_": "fastvideo.train.models.wan.WanModel", "attention_backend": "VSA_QAT_TRAIN_C256", "enable_gradient_checkpointing_type": "full", "flow_shift": 8.0, "init_from": "/mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0361-14b-c256-ode500-260922-123321/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-14B-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": 40, "num_layers": 40, "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" } }