{
  "status": "VALIDATED",
  "upstream_commit": "3196835913ce9e97b23888b32643a1207cc5ee70",
  "checkpoint": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree",
  "checkpoint_revision": "b65818d41939b5085451074fe8ca8b799f8d4921",
  "common": {
    "width": 1280,
    "height": 704,
    "num_frames": 362,
    "fps": 24,
    "sigma_grid_points": 5,
    "dit_forwards": 4,
    "seed": 1101,
    "profile": "all",
    "vsa_tile_size": 64,
    "vsa_sparsity": 0.9,
    "num_gpus": 8,
    "fsdp": true,
    "sp_size": 8,
    "tp_size": 1,
    "fa4": false,
    "vae_compile": true,
    "vae_parallel_decode": "gather",
    "source_compatibility_patch": "Only duration validation permits 362 VAE-aligned frames; applied equally to all arms."
  },
  "arms": {
    "original": [],
    "dit_mxfp8": [
      "DiT transformer block Q, K, V, O, FF input and FF output GEMMs: MXFP8 A8W8 only; retain original BF16 FSDP parameters and communication"
    ],
    "sage_attention": [
      "VSA fine sparse attention: Sage QK INT8 / PV FP8 only; original pooling, masks, coarse branch, gate and collectives"
    ],
    "vae_mxfp8": [
      "Video VAE decoder block Q, K, V, O, FF input and FF output GEMMs: MXFP8 A8W8 only"
    ],
    "vae_nvfp4": [
      "Same video VAE decoder GEMMs: NVFP4 A4W4 only"
    ],
    "dit_mxfp8_o_bf16": [
      "Relative to the page DiT MXFP8 arm: retain original BF16 attention to_out in all 50 blocks; Q/K/V and FF input/output retain the identical MXFP8 implementation."
    ],
    "dit_mxfp8_qkvo_bf16": [
      "Relative to DiT MXFP8 + BF16 O: also restore Q/K/V to original BF16. All 200 attention Q/K/V/O projections use original BF16; 100 FFN input/output projections retain the identical MXFP8 implementation."
    ],
    "dit_mxfp8_first_step_original": [
      "First DiT denoising forward: all original BF16 projections. Last three forwards: all 300 DiT Q/K/V/O and FFN projections MXFP8, identical to the existing DiT MXFP8 control. VSA kernel remains original throughout."
    ],
    "sage_first_step_original": [
      "First DiT denoising forward: original Triton VSA attention. Last three forwards: identical Sage Attention kernel to the existing Sage control. All DiT projections remain original BF16 throughout."
    ],
    "dit_mxfp8_sage_first_step_original": [
      "Forward 1: all 300 DiT Q/K/V/O and FFN projections use original BF16; original Triton VSA fine attention.",
      "Forwards 2-4: all 300 DiT projections use the existing MXFP8 implementation AND all 50 VSA fine-attention calls use the existing Sage implementation.",
      "VAE, sampling schedule, checkpoint, prompts, seed, sparse selection and communication retain the original gallery configuration."
    ]
  },
  "prompts": [
    {
      "id": "01-moon-teahouse",
      "file": "prompts/01-moon-teahouse.txt",
      "sha256": "42f62c2ecb9f90880041d51d3fa38fa6c236945ba80b6a862e080151824a50c2"
    },
    {
      "id": "02-dachshund-skatepark",
      "file": "prompts/02-dachshund-skatepark.txt",
      "sha256": "e846f0f20adfe498df2b500ac9db60e2efcfeda3cac51233e217fd407e7be5d4"
    },
    {
      "id": "03-attic-rain",
      "file": "prompts/03-attic-rain.txt",
      "sha256": "1db2772283d9ef939bc13d33a4761ca4e994ce6db28416201a8876f5a37b5745"
    },
    {
      "id": "04-doorway-worlds",
      "file": "prompts/04-doorway-worlds.txt",
      "sha256": "a5cc21ebede61549ce0ecea35afbc179cb2641786f390871e57c988f2b9a5870"
    },
    {
      "id": "05-ronin-void",
      "file": "prompts/05-ronin-void.txt",
      "sha256": "eac5ccdc6053a0e44bea5d3db9b0d5d61d772c313db0c8b24d920bd43e5e3b5d"
    },
    {
      "id": "06-anime-maid-livestream",
      "file": "prompts/06-anime-maid-livestream.txt",
      "sha256": "3122f5ba4d70acc756ddeeec87b3dd9cafdcba52bca79ae733e807ae66a248c3"
    },
    {
      "id": "07-ninja-obstacle-course",
      "file": "prompts/07-ninja-obstacle-course.txt",
      "sha256": "66f06fb2b7d7fac3ab97b6b0dae4a367df9527f1222d540801d271615c73767d"
    }
  ],
  "videos_per_prompt": 10,
  "total_videos": 70,
  "normalization": "Only add missing leading i in integrated_multimodal_description for prompts 03, 05, 07; remove user group labels; no padding, truncation or prompt rewriting.",
  "mxfp8_producer_source_sha256": "f29f83d8b5939164eaf056750807098dba9dcd20bb40db4c1d16c86ffff68862",
  "excluded_prior_optimizations": [
    "vLLM-Omni runtime",
    "QKV fusion",
    "RDMA direct kernels",
    "VSPLIT",
    "O consumer graph",
    "VAE mixed batching",
    "VAE early gather",
    "Green Contexts"
  ],
  "validated_videos": 35,
  "frozen_input_plan": "manifest.json",
  "validation_file": "validation.json",
  "extension_date": "2026-09-18",
  "extension_control": [
    "original",
    "dit_mxfp8_first_step_original",
    "sage_first_step_original"
  ],
  "extension_only_change": "Step 1: original BF16 DiT projections and original Triton VSA; steps 2-4: both existing DiT MXFP8 and Sage attention. Original VAE, checkpoint, sampling, seed, prompts and geometry retained.",
  "extension_schedule": {
    "original_forward_indices": [
      0
    ],
    "lossy_forward_indices": [
      1,
      2,
      3
    ],
    "original_forward_fraction": 0.25,
    "lossy_forward_fraction": 0.75,
    "lossy_methods": [
      "dit_mxfp8",
      "sage_attention"
    ],
    "granularity_note": "One original forward followed by three combined-lossy forwards; four total DiT forwards."
  }
}