{
  "attention": {
    "backend": "VSA_QAT_TRAIN_C128",
    "compression_gates": "trained",
    "cube_shape": [
      4,
      4,
      8
    ],
    "kind": "vsa",
    "physical_block_tokens": 64,
    "precision": {
      "coarse": "BF16",
      "compute": "BF16",
      "fine": "NVFP4 fake-quant with BF16 matmul",
      "master_weights": "FP32",
      "pv_fp8": false,
      "quant_p": true,
      "quant_v": true,
      "quantization_aware": true,
      "quantize_coarse": false,
      "quantize_fine": true,
      "two_level_p": false,
      "use_global_sf_p": false,
      "use_global_sf_qkv": false
    },
    "selection": {
      "effective_topk": "max(1, min(ceil((1-sparsity)*num_blocks), num_blocks))",
      "mode": "top-p",
      "sparsity": 0.9
    },
    "training_kernel_family": "nanovideo_triton_sparse_bf16_mm"
  },
  "checkpoint": {
    "created_at": "2026-09-20T08:15:07.968110+00:00",
    "id": "ckpt_f7d83fc1b7f04e49a0daad66a019f8fb",
    "state": "training",
    "step": 1000
  },
  "checkpoint_state": "inference_export",
  "code_provenance": {
    "commit": "95207e9c993e0a6dbe8e3d6433f885892cf5651c",
    "diff_sha256": "406397829be93e8f87ca5054a0c4e0aba61ad419549ca56a905101c3936749e8",
    "dirty": true,
    "source_archive": {
      "base_archive_sha256": "d667d04ef2e6da7896bf7f2879510cd6697d4e69d2f0958b09e9b7074d514500",
      "files_sha256": "242a24b8380e88e30617b6b52e151c390544f06b15a3b98c265e84a9ae1915f1",
      "manifest_sha256": "701d1522f39b4c8adbac593c1d04a92866d367d0538fcb1cbb79924bfafe02d8",
      "patch": "frozen_provenance_linear_qat_dense_dmd_c128.patch",
      "root": "/home/hal-ywu/git/vsqa/outputs/training-source-95207e9c-ladder-linearqat-densedmd-c128"
    },
    "untracked_sha256": {}
  },
  "config": {
    "callbacks": {
      "grad_clip": {
        "max_grad_norm": 1.0
      },
      "validation": {
        "attn_qat_infer": false,
        "dataset_file": "/home/hal-ywu/git/vsqa/logs/fastvideo-mixkit61f-vsqa-dmd_c256t8-1000-V0247-base-linfp4-mixed-b16-260919-030359/validation.json",
        "every_steps": 50,
        "guidance_scale": 1.0,
        "height": 448,
        "num_frames": 61,
        "pipeline_target": "fastvideo.pipelines.basic.wan.wan_dmd_pipeline.WanDMDPipeline",
        "run_at_start": true,
        "sampling_steps": [
          4
        ],
        "sampling_timesteps": [
          1000,
          750,
          500,
          250
        ],
        "width": 896
      }
    },
    "method": {
      "_target_": "fastvideo.train.methods.distribution_matching.dmd2.DMD2Method",
      "dmd_denoising_steps": [
        1000,
        750,
        500,
        250
      ],
      "fake_score_betas": [
        0.0,
        0.999
      ],
      "fake_score_learning_rate": 8e-06,
      "fake_score_lr_scheduler": "constant",
      "generator_update_interval": 1,
      "max_timestep_ratio": 0.98,
      "min_timestep_ratio": 0.02,
      "real_score_guidance_scale": 3.5,
      "rollout_mode": "simulate"
    },
    "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": "/home/hal-ywu/.cache/huggingface/hub/models--Wan-AI--Wan2.1-T2V-1.3B-Diffusers/snapshots/0fad780a534b6463e45facd96134c9f345acfa5b",
        "trainable": true
      },
      "student": {
        "_target_": "fastvideo.train.models.wan.WanModel",
        "attention_backend": "VSA_QAT_TRAIN_C128",
        "enable_gradient_checkpointing_type": "full",
        "flow_shift": 8.0,
        "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
        "trainable": true
      },
      "teacher": {
        "_target_": "fastvideo.train.models.wan.WanModel",
        "attention_backend": "TORCH_SDPA",
        "disable_custom_init_weights": true,
        "flow_shift": 8.0,
        "init_from": "/home/hal-ywu/.cache/huggingface/hub/models--Wan-AI--Wan2.1-T2V-1.3B-Diffusers/snapshots/0fad780a534b6463e45facd96134c9f345acfa5b",
        "trainable": false
      }
    },
    "pipeline": {
      "dit_config": {
        "quant_config": "nvfp4_qat_train"
      },
      "flow_shift": 8.0
    },
    "training": {
      "checkpoint": {
        "checkpoints_total_limit": 0,
        "output_dir": "/home/hal-ywu/git/vsqa/logs/fastvideo-mixkit61f-vsqa-dmd_c256t8-1000-V0247-base-linfp4-mixed-b16-260919-030359/checkpoints",
        "resume_from_checkpoint": "",
        "training_state_checkpointing_steps": 50
      },
      "data": {
        "data_path": "/home/hal-ywu/git/vsqa/data/mixed/training_dataset",
        "dataloader_num_workers": 0,
        "num_frames": 61,
        "num_height": 448,
        "num_latent_t": 16,
        "num_width": 896,
        "preprocessed_data_type": "text_only",
        "seed": 1000,
        "train_batch_size": 4,
        "training_cfg_rate": 0.0
      },
      "distributed": {
        "hsdp_replicate_dim": 4,
        "hsdp_shard_dim": 1,
        "num_gpus": 4,
        "sp_size": 1,
        "tp_size": 1
      },
      "dit_precision": "fp32",
      "loop": {
        "gradient_accumulation_steps": 1,
        "max_train_steps": 1000
      },
      "optimizer": {
        "betas": [
          0.0,
          0.999
        ],
        "learning_rate": 2e-06,
        "lr_scheduler": "constant",
        "lr_warmup_steps": 0,
        "weight_decay": 0.01
      },
      "tracker": {
        "entity": "alexzms-ucsd",
        "project_name": "fastvideo",
        "run_name": "fastvideo-mixkit61f-vsqa-dmd_c256t8-1000-V0247-base-linfp4-mixed-b16-260919-030359",
        "trackers": [
          "wandb"
        ]
      },
      "vsa": {
        "cache_tile_buf": false,
        "sparsity": 0.9
      }
    }
  },
  "config_sha256": {
    "resolved": "c431a59a5ba292c2868549f2d71782955ae29bfb8e9258e6a38bc7f16e0cc9d6",
    "source": "8747db4724625def9dec76dbcbdafdfcd449d0e9798591b8e661076f3431a4d8"
  },
  "metadata_schema": "fastvideo.nanovideo_vsa_training",
  "metadata_schema_version": 2,
  "model": {
    "architecture": {
      "attention_head_dim": 128,
      "in_channels": 16,
      "num_attention_heads": 12,
      "num_layers": 30,
      "out_channels": 16,
      "patch_size": [
        1,
        2,
        2
      ]
    },
    "kind": "wan-t2v-1.3b"
  },
  "provenance": {
    "code": {
      "commit": "95207e9c993e0a6dbe8e3d6433f885892cf5651c",
      "diff_sha256": "406397829be93e8f87ca5054a0c4e0aba61ad419549ca56a905101c3936749e8",
      "dirty": true,
      "source_archive": {
        "base_archive_sha256": "d667d04ef2e6da7896bf7f2879510cd6697d4e69d2f0958b09e9b7074d514500",
        "files_sha256": "242a24b8380e88e30617b6b52e151c390544f06b15a3b98c265e84a9ae1915f1",
        "manifest_sha256": "701d1522f39b4c8adbac593c1d04a92866d367d0538fcb1cbb79924bfafe02d8",
        "patch": "frozen_provenance_linear_qat_dense_dmd_c128.patch",
        "root": "/home/hal-ywu/git/vsqa/outputs/training-source-95207e9c-ladder-linearqat-densedmd-c128"
      },
      "untracked_sha256": {}
    },
    "configs": {
      "resolved": {
        "path": "resolved.yaml",
        "sha256": "c431a59a5ba292c2868549f2d71782955ae29bfb8e9258e6a38bc7f16e0cc9d6"
      },
      "source": {
        "path": "source.yaml",
        "sha256": "8747db4724625def9dec76dbcbdafdfcd449d0e9798591b8e661076f3431a4d8"
      }
    },
    "lineage": {
      "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
      "kind": "base_model_initialization"
    }
  },
  "run_identity": {
    "output_dir": "/home/hal-ywu/git/vsqa/logs/fastvideo-mixkit61f-vsqa-dmd_c256t8-1000-V0247-base-linfp4-mixed-b16-260919-030359/checkpoints",
    "tracker": {
      "entity": "alexzms-ucsd",
      "project_name": "fastvideo",
      "run_name": "fastvideo-mixkit61f-vsqa-dmd_c256t8-1000-V0247-base-linfp4-mixed-b16-260919-030359",
      "trackers": [
        "wandb"
      ]
    },
    "wandb": {
      "id": "xlonwzjx",
      "url": "https://wandb.ai/alexzms-ucsd/fastvideo/runs/xlonwzjx"
    }
  },
  "step": 1000,
  "training_contract": {
    "attention_backend": "VSA_QAT_TRAIN_C128",
    "attention_kind": "VSA",
    "cube_shape": [
      4,
      4,
      8
    ],
    "direct_lineage": {
      "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
      "kind": "base_model_initialization"
    },
    "distillation": {
      "checkpoint_roles": [
        "student",
        "critic"
      ],
      "frozen_roles": [
        "teacher"
      ],
      "method": {
        "_target_": "fastvideo.train.methods.distribution_matching.dmd2.DMD2Method",
        "dmd_denoising_steps": [
          1000,
          750,
          500,
          250
        ],
        "fake_score_betas": [
          0.0,
          0.999
        ],
        "fake_score_learning_rate": 8e-06,
        "fake_score_lr_scheduler": "constant",
        "generator_update_interval": 1,
        "max_timestep_ratio": 0.98,
        "min_timestep_ratio": 0.02,
        "real_score_guidance_scale": 3.5,
        "rollout_mode": "simulate"
      },
      "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": "/home/hal-ywu/.cache/huggingface/hub/models--Wan-AI--Wan2.1-T2V-1.3B-Diffusers/snapshots/0fad780a534b6463e45facd96134c9f345acfa5b",
          "trainable": true
        },
        "student": {
          "_target_": "fastvideo.train.models.wan.WanModel",
          "attention_backend": "VSA_QAT_TRAIN_C128",
          "enable_gradient_checkpointing_type": "full",
          "flow_shift": 8.0,
          "init_from": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
          "trainable": true
        },
        "teacher": {
          "_target_": "fastvideo.train.models.wan.WanModel",
          "attention_backend": "TORCH_SDPA",
          "disable_custom_init_weights": true,
          "flow_shift": 8.0,
          "init_from": "/home/hal-ywu/.cache/huggingface/hub/models--Wan-AI--Wan2.1-T2V-1.3B-Diffusers/snapshots/0fad780a534b6463e45facd96134c9f345acfa5b",
          "trainable": false
        }
      }
    },
    "kernel_source_commit": "e8b4630e0ee15be13d1e0f6823882b776edbe62e",
    "kernel_source_files": {
      "block_sparse_attn.py": "da77938f5e68822e57b7571507064199c62440d580efa9e3b14f9b5d1bc72199",
      "qat/nvfp4_utils.py": "70b613fca22817ca764496095efc73127515574a53e4b9fcc70bd56a63b3ac3e",
      "qat/quant_utils.py": "e450efca2c375d48889c9d647f5760c9fe90330436dfc623a5684ebe6a48080d",
      "qat/sparse/bf16_mm.py": "1c8bbf7ac573ccfa3d28b41d001c922cb261c1c74bcc8e83743e0da092c82ff5",
      "triton_kernels/block_sparse_attn.py": "715d4d51ff26f3edf9764b6708769376068f228bedd1f7db26aee93375203d83",
      "triton_kernels/index.py": "ce8beccf1f66554917765976820834fdf03a45ad0d1852e6da0b16c811d7b19d"
    },
    "linear_quantization": {
      "activation_global_scale": "dynamic, 448*6/max|X| recomputed every forward",
      "backward": "whole-linear STE on the saved unquantized X and master W in the gradient dtype",
      "format": "NVFP4 W4A4: E2M1 values, 16-element blocks, FP8 E4M3 block scales, FP32 global scale",
      "forward": "flashinfer nvfp4_quantize + mm_fp4 (cutlass) real FP4 GEMM, BF16 output",
      "kind": "nvfp4_qat_train",
      "master_weights": "FP32",
      "scope": "student transformer only; teacher and critic load with quant_config dropped",
      "target_prefix_substrings": [
        "ffn.fc_in",
        "ffn.fc_out",
        "to_q",
        "to_k",
        "to_v",
        "to_out",
        "self_attention.out_layer",
        "cross_attention.out_layer",
        "feed_forward.mlp.fc_in",
        "feed_forward.mlp.fc_out"
      ],
      "weight_global_scale": "dynamic, 448*6/max|W| recomputed every forward from the master weight"
    },
    "model_architecture": {
      "attention_head_dim": 128,
      "in_channels": 16,
      "num_attention_heads": 12,
      "num_layers": 30,
      "out_channels": 16,
      "patch_size": [
        1,
        2,
        2
      ]
    },
    "model_kind": "Wan bidirectional T2V",
    "physical_block_tokens": 64,
    "precision": {
      "coarse": "BF16",
      "compute": "BF16",
      "fine": "NVFP4 fake-quant with BF16 matmul",
      "master_weights": "FP32",
      "pv_fp8": false,
      "quant_p": true,
      "quant_v": true,
      "quantize_coarse": false,
      "quantize_fine": true,
      "two_level_p": false,
      "use_global_sf_p": false,
      "use_global_sf_qkv": false
    },
    "selection": {
      "effective_topk": "max(1, min(ceil((1-sparsity)*num_blocks), num_blocks))",
      "mode": "top-p",
      "sparsity": 0.9
    },
    "training_kernel_family": "nanovideo_triton_sparse_bf16_mm"
  }
}