{
  "date": "2026-09-23",
  "request": {
    "task": "t2va",
    "width": 1344,
    "height": 768,
    "frames": 124,
    "fps": 24,
    "seed": 1101,
    "dit_forwards": 4,
    "prompt_sha256": "5abb96b31333eea455c241aa17e61d1eaf75d7e78625c80f17302b34ce11fad3"
  },
  "timing_boundary": "Local HTTP submission to complete MP4 response bytes; excludes model loading, two warmups and post-response validation.",
  "protocol": "Five sequential requests per arm. Publish repeat 1 without re-encoding. Report actual decoded frame count. Identical request, seed and prompt across arms.",
  "arms": [
    {
      "key": "optimized-base",
      "label": "Optimized baseline",
      "track": "optimized",
      "request_seconds": [
        4.666182689135894,
        4.675563522148877,
        4.681992647936568,
        4.690105772810057,
        4.698794882046059
      ],
      "median_seconds": 4.681992647936568,
      "min_seconds": 4.666182689135894,
      "max_seconds": 4.698794882046059,
      "median_rtf": 0.9061920669430732,
      "under_5": 5,
      "under_playback": 5,
      "warmup_seconds": [
        16.340221380116418,
        4.692218508105725
      ],
      "load_seconds": 422.25831137411296,
      "published_repeat": 1,
      "media": {
        "width": 1344,
        "height": 768,
        "frames": 124,
        "fps": 24,
        "video_seconds": 5.166667,
        "audio_seconds": 5.175,
        "audio_rate": 32000,
        "audio_channels": 2,
        "bytes": 10206291,
        "sha256": "d4901bef7a356fdfc4cc1505629988e4d993db6744f61bf7aeb876ad126f35bd",
        "decode": "PASS"
      },
      "asset_prefix": "benchmarks/ulysses-5s-768p",
      "encoder": {
        "video": "H.264 / libx264",
        "audio": "AAC",
        "crf": 18.0
      }
    },
    {
      "key": "optimized-fences",
      "label": "Optimized + PR #5317 fences",
      "track": "optimized",
      "request_seconds": [
        4.659823487047106,
        4.66123240115121,
        4.67330171703361,
        4.682930525857955,
        4.69424985605292
      ],
      "median_seconds": 4.67330171703361,
      "min_seconds": 4.659823487047106,
      "max_seconds": 4.69424985605292,
      "median_rtf": 0.9045099513929598,
      "under_5": 5,
      "under_playback": 5,
      "warmup_seconds": [
        22.814474185928702,
        4.6871386668644845
      ],
      "load_seconds": 144.0978217869997,
      "published_repeat": 1,
      "media": {
        "width": 1344,
        "height": 768,
        "frames": 124,
        "fps": 24,
        "video_seconds": 5.166667,
        "audio_seconds": 5.175,
        "audio_rate": 32000,
        "audio_channels": 2,
        "bytes": 10206291,
        "sha256": "d4901bef7a356fdfc4cc1505629988e4d993db6744f61bf7aeb876ad126f35bd",
        "decode": "PASS"
      },
      "asset_prefix": "benchmarks/ulysses-5s-768p",
      "encoder": {
        "video": "H.264 / libx264",
        "audio": "AAC",
        "crf": 18.0
      }
    },
    {
      "key": "original",
      "label": "Original FastH3",
      "track": "original",
      "request_seconds": [
        22.814603788079694,
        23.259365485049784,
        22.888991127023473,
        22.91032107290812,
        22.834863323019817
      ],
      "median_seconds": 22.888991127023473,
      "min_seconds": 22.814603788079694,
      "max_seconds": 23.259365485049784,
      "median_rtf": 4.430127029093121,
      "under_5": 0,
      "under_playback": 0,
      "warmup_seconds": [
        42.00765839894302,
        23.166222010040656
      ],
      "load_seconds": 482.1820451640524,
      "published_repeat": 1,
      "media": {
        "width": 1344,
        "height": 768,
        "frames": 124,
        "fps": 24,
        "video_seconds": 5.166667,
        "audio_seconds": 5.152,
        "audio_rate": 32000,
        "audio_channels": 2,
        "bytes": 5369058,
        "sha256": "fd3538b15923cc517dcfb19e7128edd7bc0743e69889e8cff89a8b78fafe93e7",
        "decode": "PASS"
      },
      "asset_prefix": "benchmarks/ulysses-5s-768p",
      "encoder": {
        "video": "H.264 / libx264",
        "audio": "AAC",
        "crf": 23.0
      },
      "denoise_seconds": [
        11.244829176925123,
        11.2813191160094,
        11.199471672065556,
        11.24797152588144,
        11.204126084921882
      ],
      "stage_wall_seconds": {
        "input_preparation_stage": [
          4.5428983867168427e-05,
          8.332799188792706e-05,
          5.211890675127506e-05,
          5.588890053331852e-05,
          6.938795559108257e-05
        ],
        "conditioning_stage": [
          1.8788065649569035,
          1.905230934964493,
          1.9118711261544377,
          1.906802725046873,
          1.9111264490056783
        ],
        "latent_preparation_stage": [
          0.009790165815502405,
          0.009703906951472163,
          0.01492076413705945,
          0.012089625000953674,
          0.015323068015277386
        ],
        "denoising_stage": [
          11.244829176925123,
          11.2813191160094,
          11.199471672065556,
          11.24797152588144,
          11.204126084921882
        ],
        "video_decoding_stage": [
          7.313360863132402,
          7.600872442126274,
          7.356753197964281,
          7.322395568946376,
          7.2910798660013825
        ],
        "audio_decoding_stage": [
          0.3981396700255573,
          0.39556411700323224,
          0.4254618939012289,
          0.38724817405454814,
          0.3885223811957985
        ],
        "PostDecodeFrameProcessStage": [
          0.6207391561474651,
          0.6190282572060823,
          0.6133234580047429,
          0.6357368589378893,
          0.6456413629930466
        ],
        "VideoSaveStage": [
          0.11881457408890128,
          0.16358501696959138,
          0.1354449880309403,
          0.12236602092161775,
          0.13855870300903916
        ],
        "AudioMuxStage": [
          0.0,
          0.0,
          0.0,
          0.0,
          0.0
        ]
      }
    },
    {
      "key": "original-workspace",
      "label": "Original + PR #5027 workspace",
      "track": "original",
      "request_seconds": [
        23.058526334818453,
        22.790835798950866,
        22.63426429196261,
        22.44739115308039,
        22.6047210409306
      ],
      "median_seconds": 22.63426429196261,
      "min_seconds": 22.44739115308039,
      "max_seconds": 23.058526334818453,
      "median_rtf": 4.380825064197597,
      "under_5": 0,
      "under_playback": 0,
      "warmup_seconds": [
        43.70412983512506,
        23.18850973295048
      ],
      "load_seconds": 444.1533938369248,
      "published_repeat": 1,
      "media": {
        "width": 1344,
        "height": 768,
        "frames": 124,
        "fps": 24,
        "video_seconds": 5.166667,
        "audio_seconds": 5.152,
        "audio_rate": 32000,
        "audio_channels": 2,
        "bytes": 5369555,
        "sha256": "a61487f494c8e2b897a20819853894d4eaf4e518c239311132cf9f918e946fa2",
        "decode": "PASS"
      },
      "asset_prefix": "benchmarks/ulysses-5s-768p",
      "encoder": {
        "video": "H.264 / libx264",
        "audio": "AAC",
        "crf": 23.0
      },
      "denoise_seconds": [
        11.199054844211787,
        11.160045218886808,
        11.150640819221735,
        11.232963623013347,
        11.153900409117341
      ],
      "stage_wall_seconds": {
        "input_preparation_stage": [
          6.748898886144161e-05,
          0.00011590821668505669,
          4.922901280224323e-05,
          4.270905628800392e-05,
          6.536906585097313e-05
        ],
        "conditioning_stage": [
          1.8957117991521955,
          1.9026620739605278,
          1.9205130559857935,
          1.860743235098198,
          1.84036171506159
        ],
        "latent_preparation_stage": [
          0.012913760961964726,
          0.011189450975507498,
          0.016332879895344377,
          0.009142766939476132,
          0.014526073122397065
        ],
        "denoising_stage": [
          11.199054844211787,
          11.160045218886808,
          11.150640819221735,
          11.232963623013347,
          11.153900409117341
        ],
        "video_decoding_stage": [
          7.378072980092838,
          7.399627143982798,
          7.130414603976533,
          7.008375684032217,
          7.225206404924393
        ],
        "audio_decoding_stage": [
          0.42509215301834047,
          0.4159305449575186,
          0.4322905149310827,
          0.3981144009158015,
          0.4314662991091609
        ],
        "PostDecodeFrameProcessStage": [
          0.6182961470913142,
          0.6184814930893481,
          0.6195139749906957,
          0.6437355049420148,
          0.6168057119939476
        ],
        "VideoSaveStage": [
          0.11754485499113798,
          0.1108270650729537,
          0.12082758685573936,
          0.12534924712963402,
          0.1208922560326755
        ],
        "AudioMuxStage": [
          0.0,
          0.0,
          0.0,
          0.0,
          0.0
        ]
      }
    }
  ],
  "pr5027": {
    "head": "e4a8e35059235306908b32a8186b958586828a70",
    "tested_subset": "Reusable NCCL workspace and explicit out buffers for stock FastVideo serialized Ulysses. Original Triton sparse attention and gate retained.",
    "not_tested": "Head-chunk attention pipeline; not directly compatible with the optimized custom PCIe/RDMA producer-direct API."
  },
  "pr5317": {
    "head": "b1f3a95a7fcec2af73bf520dcf06ae8aff6fa6d1",
    "tested_subset": "K/V shared-memory proxy fences backported to existing BF16 and Sage SM120 attention kernels; end-to-end optimized arm uses Sage.",
    "not_tested": "SM90/SM100-specific attention paths, FP8 wire transport and full native head-chunk pipeline."
  },
  "scope": "Two separate within-stack PR comparisons. The stock full checkpoint and optimized base-plus-adapter deployment are different artifacts; cross-stack differences are not attributed solely to PRs or precision.",
  "provenance": {
    "original_source": "FastVideo 3196835913ce9e97b23888b32643a1207cc5ee70, existing gallery snapshot",
    "original_checkpoint_revision": "b65818d41939b5085451074fe8ca8b799f8d4921",
    "original_torch": "2.13.0",
    "optimized_torch": "2.13.0+cu132",
    "cutlass_dsl": "4.6.2",
    "triton": "3.7.1",
    "original_ulysses_rows": 38168,
    "optimized_ulysses_rows": 38208
  }
}