{
  "schema_version": 1,
  "id": "r3-deployment-qk-fc2-native-norm",
  "campaign_id": "h3-m5pro-round-03",
  "contract_sha256": "cbd4ce596524b59629216716299ee8677b10f49140de962384153aaaa2e55dec",
  "complete": true,
  "change_class": "exact",
  "hypothesis": "Native reduction matching QK/RoPE fusion and shape-qualified Metal 4 FC2 tiling reduce full generation latency without changing a denoising tensor or pixel.",
  "context": {
    "hardware": "Apple M5 Pro, 16 GPU cores, 64 GiB unified memory",
    "os": "macOS 27.0.1 arm64",
    "software": "MLX/MLX Metal 0.32.4; Python 3.12.14; strict precision environment",
    "environment": "AC Automatic power; MLX_ENABLE_TF32=0, MLX_CONV_WINOGRAD=0; user background apps unchanged; no thermal warning recorded",
    "timing_boundary": "Pipeline.generate total_s: warm text cache, original CPU RNG, six heads, decode, trace and PNG save. Excludes interpreter startup and external quality scoring.",
    "memory_metric": "Parent macOS physical footprint sampled at 100ms, current-interval maximum; includes Metal, excludes compiler XPC. MLX allocator separately; do not sum. Both compiled variants may be resident.",
    "cache_state": "Resident shared native weights and VAE; shape warmup for each variant. Exact conditioning caches warmed outside timed population.",
    "execution_policy": "Single GPU benchmark process; per-case repeats alternate control/candidate then candidate/control then control/candidate. Warmup and first-load excluded from timed population."
  },
  "variant": {
    "id": "qk-fc2-native-norm",
    "code_sha256": "8121aef031c28e2af6aa47e2492040619695b465eca661413576fcc00e6e58ea",
    "model_sha256": "3274a3e5fe0d2ea0d63f66b020beb193276a0a9d8de7ea466b312fbae9a2746f",
    "config_sha256": "d8ede93ee4766952aa052d0f550a675508f271dcd04c064747713fdf8c4cc24e",
    "command": [
      ".venv/bin/python",
      "-m",
      "experiments.r3.images",
      "--deployment",
      "--variants",
      "qk-fc2-native-norm",
      "--workload",
      "optimization/round-03/workload.json",
      "--model",
      "/Users/devin/Desktop/minimax-h3-image-opt/models/MiniMax-H3-Image",
      "--text-encoder",
      "/Users/devin/Desktop/minimax-h3-image-opt/models/MiniMax-H3/text_encoder",
      "--output",
      "outputs/optimization-r3-qk"
    ]
  },
  "quality": {
    "variant_sha256": "11f7dd242ba190be2c0fd67530140cd7da6154960cdd902126b7ebb3a56f3ec6",
    "reference_sha256": "201be3382041aec890cbd766c2f88a43799f7e2d33807e9a1504f7e29e422065",
    "evaluator_sha256": "4db0c5e8606ffc011bf85dbadabc482468a2e49cd5a1757a774620d493eba7ee",
    "outputs": {
      "product": "eaed6c290ea3f7843ed4a789614cf3ad384d5acf3d0e0cc40a569fdcb850390f",
      "portrait": "4226310bd1d79a0d0d4c0b572e2e7729541f5f171260cfeefd5e73b0a4196a47",
      "edit": "b4b5875fea6451b9c7c24efc7903945a454ae7885ed3016bfc361b4799b2508e",
      "holdout-detail": "561b18aa708d61a5accbfd12f2280e8f7e303505324c43471ba06d60fe0351d5"
    },
    "metrics": {
      "max_abs_error": 0,
      "nonfinite_values": 0
    }
  },
  "runs": [
    {
      "case": "product",
      "repeat": 0,
      "seconds": 63.90098704188131,
      "peak_memory_bytes": 42450606784,
      "output_sha256": "eaed6c290ea3f7843ed4a789614cf3ad384d5acf3d0e0cc40a569fdcb850390f"
    },
    {
      "case": "product",
      "repeat": 1,
      "seconds": 64.23788916692138,
      "peak_memory_bytes": 42492910320,
      "output_sha256": "eaed6c290ea3f7843ed4a789614cf3ad384d5acf3d0e0cc40a569fdcb850390f"
    },
    {
      "case": "product",
      "repeat": 2,
      "seconds": 61.27363566705026,
      "peak_memory_bytes": 42598718264,
      "output_sha256": "eaed6c290ea3f7843ed4a789614cf3ad384d5acf3d0e0cc40a569fdcb850390f"
    },
    {
      "case": "portrait",
      "repeat": 0,
      "seconds": 60.65490966709331,
      "peak_memory_bytes": 42673544112,
      "output_sha256": "4226310bd1d79a0d0d4c0b572e2e7729541f5f171260cfeefd5e73b0a4196a47"
    },
    {
      "case": "portrait",
      "repeat": 1,
      "seconds": 61.10937987500802,
      "peak_memory_bytes": 42671692696,
      "output_sha256": "4226310bd1d79a0d0d4c0b572e2e7729541f5f171260cfeefd5e73b0a4196a47"
    },
    {
      "case": "portrait",
      "repeat": 2,
      "seconds": 62.00676687504165,
      "peak_memory_bytes": 42665237448,
      "output_sha256": "4226310bd1d79a0d0d4c0b572e2e7729541f5f171260cfeefd5e73b0a4196a47"
    }
  ],
  "evidence": [
    {
      "path": "outputs/optimization-r3-qk/runs.jsonl",
      "sha256": "ec303dd7f11fa79c1539a424d04ff9dc9f427a3d4181a7957199b7dc6ee6b9c8"
    },
    {
      "path": "outputs/optimization-r3-qk/source-manifest.json",
      "sha256": "8121aef031c28e2af6aa47e2492040619695b465eca661413576fcc00e6e58ea"
    },
    {
      "path": "reports/optimization-r3/qk-deployment/quality-audit.json",
      "sha256": "3d242d63e283fa6dc645adee2c3a1dfda59b5d9ae4ac56cf04e15a5486816a2e"
    },
    {
      "path": "reports/optimization-r3/qk-deployment/summary.json",
      "sha256": "4dcff0c7944ac8d80a9c9961a1a90507e465ee1b610f86441a45804e02cc3cbe"
    }
  ],
  "extensions": {
    "canonical_output": "SHA256 decoded RGB uint8 bytes in row-major order",
    "model_identity": "model_sha256 is the pinned model/source provenance manifest hash; no new full checkpoint rehash",
    "reference_scope": "Incremental identity to MLX BF16 recipe, not CUDA or unpublished FP32 training weights",
    "trace_identical": true
  }
}