{
  "schema_version": 1,
  "source_of_truth": "cmp/docs/results/*.md and cmp/patches/manifest.json#inherited_results",
  "note": "Reviewed public measurements and inherited patch records; see the linked pages for interpretation.",
  "entry_count": 8,
  "entries": [
    {
      "mechanism": "sm70-d256-shared-q",
      "title": "SM70 D256 shared Q and unit rescaling preserve output through 250K",
      "date": "2026-09-07",
      "verdict": "kept",
      "selectors": [],
      "contexts": [
        8192,
        65536,
        131072,
        250000
      ],
      "byte_exact": true,
      "deltas": [
        {
          "metric": "request wall time",
          "context": 8192,
          "change_pct": -0.316226,
          "model": "primary",
          "devices": 6,
          "note": "Shared Q plus unit rescaling against saved original references; single samples and uncontrolled drift."
        },
        {
          "metric": "prompt processing time",
          "context": 8192,
          "change_pct": -0.417361,
          "model": "primary",
          "devices": 6,
          "note": "Shared Q plus unit rescaling against saved original references; single samples and uncontrolled drift."
        },
        {
          "metric": "generation time",
          "context": 8192,
          "change_pct": 0.081463,
          "model": "primary",
          "devices": 6,
          "note": "Shared Q plus unit rescaling against saved original references; single samples and uncontrolled drift."
        },
        {
          "metric": "request wall time",
          "context": 65536,
          "change_pct": -1.386239,
          "model": "primary",
          "devices": 6,
          "note": "Shared Q plus unit rescaling against saved original references; single samples and uncontrolled drift."
        },
        {
          "metric": "prompt processing time",
          "context": 65536,
          "change_pct": -1.510815,
          "model": "primary",
          "devices": 6,
          "note": "Shared Q plus unit rescaling against saved original references; single samples and uncontrolled drift."
        },
        {
          "metric": "generation time",
          "context": 65536,
          "change_pct": 2.482593,
          "model": "primary",
          "devices": 6,
          "note": "Shared Q plus unit rescaling against saved original references; single samples and uncontrolled drift."
        },
        {
          "metric": "request wall time",
          "context": 131072,
          "change_pct": -2.161571,
          "model": "primary",
          "devices": 6,
          "note": "Shared Q plus unit rescaling against saved original references; single samples and uncontrolled drift."
        },
        {
          "metric": "prompt processing time",
          "context": 131072,
          "change_pct": -2.183743,
          "model": "primary",
          "devices": 6,
          "note": "Shared Q plus unit rescaling against saved original references; single samples and uncontrolled drift."
        },
        {
          "metric": "generation time",
          "context": 131072,
          "change_pct": -0.917041,
          "model": "primary",
          "devices": 6,
          "note": "Shared Q plus unit rescaling against saved original references; single samples and uncontrolled drift."
        },
        {
          "metric": "request wall time",
          "context": 250000,
          "change_pct": -2.676602,
          "model": "primary",
          "devices": 6,
          "note": "Shared Q plus unit rescaling against saved original references; single samples and uncontrolled drift.",
          "headline": true
        },
        {
          "metric": "prompt processing time",
          "context": 250000,
          "change_pct": -2.710504,
          "model": "primary",
          "devices": 6,
          "note": "Shared Q plus unit rescaling against saved original references; single samples and uncontrolled drift."
        },
        {
          "metric": "generation time",
          "context": 250000,
          "change_pct": -0.635263,
          "model": "primary",
          "devices": 6,
          "note": "Shared Q plus unit rescaling against saved original references; single samples and uncontrolled drift."
        },
        {
          "metric": "unit-scale request wall time vs original reference",
          "context": 8192,
          "change_pct": -0.316226,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale request wall time vs first shared-Q",
          "context": 8192,
          "change_pct": 0.064568,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale prompt processing time vs original reference",
          "context": 8192,
          "change_pct": -0.417361,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale prompt processing time vs first shared-Q",
          "context": 8192,
          "change_pct": -0.048316,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale generation time vs original reference",
          "context": 8192,
          "change_pct": 0.081463,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale generation time vs first shared-Q",
          "context": 8192,
          "change_pct": 0.484038,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale request wall time vs original reference",
          "context": 65536,
          "change_pct": -1.386239,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale request wall time vs first shared-Q",
          "context": 65536,
          "change_pct": -0.030149,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale prompt processing time vs original reference",
          "context": 65536,
          "change_pct": -1.510815,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale prompt processing time vs first shared-Q",
          "context": 65536,
          "change_pct": -0.076213,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale generation time vs original reference",
          "context": 65536,
          "change_pct": 2.482593,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale generation time vs first shared-Q",
          "context": 65536,
          "change_pct": 1.383658,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale request wall time vs original reference",
          "context": 131072,
          "change_pct": -2.161571,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale request wall time vs first shared-Q",
          "context": 131072,
          "change_pct": 0.024788,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale prompt processing time vs original reference",
          "context": 131072,
          "change_pct": -2.183743,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale prompt processing time vs first shared-Q",
          "context": 131072,
          "change_pct": 0.011567,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale generation time vs original reference",
          "context": 131072,
          "change_pct": -0.917041,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale generation time vs first shared-Q",
          "context": 131072,
          "change_pct": 0.784755,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale request wall time vs original reference",
          "context": 250000,
          "change_pct": -2.676602,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale request wall time vs first shared-Q",
          "context": 250000,
          "change_pct": -0.073207,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale prompt processing time vs original reference",
          "context": 250000,
          "change_pct": -2.710504,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale prompt processing time vs first shared-Q",
          "context": 250000,
          "change_pct": -0.087055,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale generation time vs original reference",
          "context": 250000,
          "change_pct": -0.635263,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        },
        {
          "metric": "unit-scale generation time vs first shared-Q",
          "context": 250000,
          "change_pct": 0.762114,
          "model": "primary",
          "devices": 6,
          "note": "Single sample; saved comparison; uncontrolled drift."
        }
      ],
      "commit": "4fe5ecd7536bad00fa4f589a293e9284d2086302",
      "upstream_base": "4d9176092d00586775af140581bb0b558ddc4389",
      "upstream_prs": [],
      "patch": "cmp/patches/0008-sm70-d256-shared-q.patch",
      "evidence": "cmp/docs/results/data/sm70-d256-shared-q-promotion.json",
      "promotion": "accepted",
      "tags": [
        "d256",
        "flash-attention",
        "research-snapshot",
        "shared-memory",
        "sm70"
      ],
      "supersedes": [],
      "superseded_by": null,
      "data": [
        "sm70-d256-shared-q-20260907.csv",
        "sm70-d256-shared-q-unroll4-20260907.csv",
        "sm70-d256-shared-q-unit-scale-20260907.csv",
        "sm70-d256-shared-q-promotion.csv"
      ],
      "template": false,
      "source": "cmp/docs/results/2026-09-07-sm70-d256-shared-q.md",
      "url": "results/2026-09-07-sm70-d256-shared-q/",
      "summary": "The combined shared-Q and unit-rescale patch was accepted after review of its gains, regressions and measurement limits. All four primary contexts preserved exact 256-token output. At 250K, observed request wall time was 2.677% lower than the original saved reference. These single samples do not establish statistical repeatability or native kernel attribution."
    },
    {
      "mechanism": "dp4a-mmq-routing",
      "title": "DP4A MMQ routing",
      "date": "2026-09-05",
      "verdict": "kept",
      "source_patch": "cmp/patches/0001-dp4a-mmq-routing.patch",
      "commit": "4e66a8481b33f41e8f25b64d083059c529a168a3",
      "selectors": [
        "--cuda-mmq force"
      ],
      "contexts": [
        128000
      ],
      "byte_exact": true,
      "headline": "2.36x matched prefill throughput at 128K in one six-card configuration; not a decode claim."
    },
    {
      "mechanism": "volta-q4k-q5k-dp4a",
      "title": "Volta Q4_K and Q5_K DP4A",
      "date": "2026-09-05",
      "verdict": "kept",
      "source_patch": "cmp/patches/0002-volta-q4k-q5k-dp4a.patch",
      "commit": "703d046c78b3ba7565f2605d54e7ba7f701afce6",
      "selectors": [
        "--cuda-mmq force"
      ],
      "contexts": [
        4096,
        100000
      ],
      "byte_exact": true,
      "headline": "+33.7% at a 4K prompt and +0.09% at an exact 100K prompt."
    },
    {
      "mechanism": "parallel-model-load",
      "title": "Parallel model load",
      "date": "2026-09-05",
      "verdict": "kept",
      "source_patch": "cmp/patches/0003-parallel-model-load.patch",
      "commit": "bc5cf34c680c898d2c7d9a974c848086df762024",
      "selectors": [
        "LLAMA_MODEL_LOAD_PARALLEL=1"
      ],
      "contexts": [],
      "byte_exact": true,
      "headline": "51-59% faster matched cold-start readiness; not token-generation throughput."
    },
    {
      "mechanism": "mapped-host-bridge",
      "title": "Mapped pinned host bridge",
      "date": "2026-09-05",
      "verdict": "kept",
      "source_patch": "cmp/patches/0004-mapped-host-bridge.patch",
      "commit": "f9859b80a39c5a133612363538f8a343d7f4a447",
      "selectors": [
        "GGML_CUDA_MAPPED_HOST_BRIDGE=1"
      ],
      "contexts": [],
      "byte_exact": true,
      "headline": "Four lossless mapped-pinned staging slots for cross-card boundaries up to 64 MiB."
    },
    {
      "mechanism": "research-selectors",
      "title": "Research selector registry",
      "date": "2026-09-05",
      "verdict": "kept",
      "source_patch": "cmp/patches/0005-research-selectors.patch",
      "commit": "e5e027c3fcf638749a39438e4d4e7c9e93a994ee",
      "selectors": [],
      "contexts": [],
      "byte_exact": true,
      "headline": "Runtime selector-read reporting; no behavior change by itself."
    },
    {
      "mechanism": "prompt-graph-capture-seed",
      "title": "Prompt graph capture seed",
      "date": "2026-09-05",
      "verdict": "kept",
      "source_patch": "cmp/patches/0006-prompt-graph-capture-seed.patch",
      "commit": "3f6f7ee022644bf913987f37ca683edbdb6c464e",
      "selectors": [
        "GGML_CUDA_PROMPT_GRAPH_CAPTURE_SEED=1"
      ],
      "contexts": [],
      "byte_exact": true,
      "headline": "Model-dependent prompt-throughput effect; first observation is captured but not replayed."
    },
    {
      "mechanism": "device-side-mask-expansion",
      "title": "Device-side mask expansion",
      "date": "2026-09-05",
      "verdict": "kept",
      "source_patch": "cmp/patches/0007-device-side-mask-expansion.patch",
      "commit": "636415e0f216110dfab887bc81cfa9e324177c70",
      "selectors": [],
      "contexts": [],
      "byte_exact": true,
      "headline": "Removes repeated expanded-mask uploads for causal masks and sparse-attention block bias."
    }
  ],
  "by_selector": {
    "--cuda-mmq force": [
      "dp4a-mmq-routing",
      "volta-q4k-q5k-dp4a"
    ],
    "GGML_CUDA_MAPPED_HOST_BRIDGE=1": [
      "mapped-host-bridge"
    ],
    "GGML_CUDA_PROMPT_GRAPH_CAPTURE_SEED=1": [
      "prompt-graph-capture-seed"
    ],
    "LLAMA_MODEL_LOAD_PARALLEL=1": [
      "parallel-model-load"
    ]
  },
  "by_verdict": {
    "kept": [
      "device-side-mask-expansion",
      "dp4a-mmq-routing",
      "mapped-host-bridge",
      "parallel-model-load",
      "prompt-graph-capture-seed",
      "research-selectors",
      "sm70-d256-shared-q",
      "volta-q4k-q5k-dp4a"
    ]
  },
  "by_tag": {
    "d256": [
      "sm70-d256-shared-q"
    ],
    "flash-attention": [
      "sm70-d256-shared-q"
    ],
    "research-snapshot": [
      "sm70-d256-shared-q"
    ],
    "shared-memory": [
      "sm70-d256-shared-q"
    ],
    "sm70": [
      "sm70-d256-shared-q"
    ]
  }
}