{
  "date": "2026-05-06",
  "model": {
    "hfId": "Qwen/Qwen3.6-27B",
    "localPath": "/home/steve/models/qwen3.6-27b-q4_0-gguf/Qwen3.6-27B-Q4_0.gguf",
    "quantization": "Q4_0"
  },
  "hardware": {
    "gpu": "Intel Arc Pro B70",
    "vramGbPerCard": 32,
    "cpu": "AMD EPYC 9015 8-Core Processor",
    "ramGb": 16,
    "os": "Ubuntu 24.04.4 LTS",
    "kernel": "6.17.0-23-generic"
  },
  "software": {
    "engine": "llama.cpp",
    "commit": "db44417",
    "backend": "SYCL Level Zero",
    "build": "/home/steve/src/llama.cpp-q4-b70/build-sycl-2026-bmg-g31/bin/llama-bench",
    "newGates": [
      "GGML_SYCL_FUSE_MMVQ2",
      "GGML_SYCL_FUSE_MMVQ2_PROBE"
    ],
    "qualityPreserving": "single-card PPL matched; 3-card forced non-conversation generation matched; 3-card PPL harness is nondeterministic even with fusion disabled, so multi-GPU quality is not fully cleared",
    "powerLimitChanges": false
  },
  "localmaxxing": {
    "submitted": true,
    "id": "cmotg6zz50004jo04eyqocxli",
    "note": "Accepted with reduced payload because full engineFlags payload returned HTTP 500 and backend enum rejected sycl-level-zero."
  },
  "graphProbe": {
    "singleB70": {
      "selector": "level_zero:2",
      "promptTokens": 0,
      "outputTokens": 1,
      "candidatePairsPerGraph": 80,
      "ffnGateUpPairs": 64,
      "vcurKcurPairs": 16,
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-fusemmvq2-probe-single2-p0n1-r1-20260506T020938Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-fusemmvq2-probe-single2-p0n1-r1-20260506T020938Z.log"
    },
    "tripleB70": {
      "selector": "level_zero:2,1,3",
      "promptTokens": 0,
      "outputTokens": 1,
      "candidatePairHitsAcrossTwoExecutions": 480,
      "normalizedModelLevelPairsPerToken": 80,
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-fusemmvq2-probe-triple213-p0n1-r1-20260506T021019Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-fusemmvq2-probe-triple213-p0n1-r1-20260506T021019Z.log"
    }
  },
  "prototypeQ8Esimd": {
    "source": "/home/steve/llm-optimization-artifacts/prototypes/q4_0_esimd_probe/bench_q4_0_q8_1_esimd.cpp",
    "binary": "/home/steve/llm-optimization-artifacts/prototypes/q4_0_esimd_probe/bench_q4_0_q8_1_esimd",
    "buildCommand": "source /opt/intel/oneapi/setvars.sh && icpx -O3 -std=c++20 -fsycl -fsycl-device-code-split=per_kernel -fsycl-targets=spir64_gen -Xs \"-device bmg\" bench_q4_0_q8_1_esimd.cpp -o bench_q4_0_q8_1_esimd",
    "results": [
      {"n": 17408, "k": 5120, "singleMedianUs": 127.916, "fused2MedianUs": 191.146, "fused2VsTwoSinglesSpeedupPct": 25.3},
      {"n": 5808, "k": 5120, "singleMedianUs": 48.854, "fused2MedianUs": 59.896, "fused2VsTwoSinglesSpeedupPct": 38.7},
      {"n": 4352, "k": 5120, "singleMedianUs": 39.271, "fused2MedianUs": 55.521, "fused2VsTwoSinglesSpeedupPct": 29.3},
      {"n": 1024, "k": 5120, "singleMedianUs": 11.146, "fused2MedianUs": 13.750, "fused2VsTwoSinglesSpeedupPct": 38.3},
      {"n": 5120, "k": 6144, "singleMedianUs": 50.625, "fused2MedianUs": 72.084, "fused2VsTwoSinglesSpeedupPct": 28.8}
    ]
  },
  "inTreeBenchmarks": {
    "singleB70Short": {
      "selector": "level_zero:2",
      "promptTokens": 0,
      "outputTokens": 128,
      "repetitions": 2,
      "controlTokSOut": 24.340948,
      "fusedTokSOut": 24.600702
    },
    "tripleB70Short": {
      "selector": "level_zero:2,1,3",
      "promptTokens": 0,
      "outputTokens": 128,
      "repetitions": 2,
      "controlTokSOut": 43.789841,
      "fusedTokSOut": 45.322169
    },
    "tripleB70Full": {
      "selector": "level_zero:2,1,3",
      "promptTokens": 512,
      "outputTokens": 512,
      "repetitions": 3,
      "promptTokS": 135.810565,
      "decodeTokS": 46.117650,
      "computedTotalTokS": 68.854236,
      "localmaxxingId": "cmotg6zz50004jo04eyqocxli",
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-fusemmvq2-fused-triple213-p512n512-r3-20260506T023132Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-fusemmvq2-fused-triple213-p512n512-r3-20260506T023132Z.log"
    },
    "quadB70Short": {
      "selector": "level_zero:0,1,2,3",
      "promptTokens": 0,
      "outputTokens": 128,
      "repetitions": 2,
      "fusedTokSOut": 34.842442
    }
  },
  "correctness": {
    "singleB70Ppl": {
      "selector": "level_zero:2",
      "batchSize": 1,
      "ubatchSize": 1,
      "contextLength": 64,
      "chunks": 1,
      "controlPpl": "2.0500 +/- 0.56981",
      "fusedPpl": "2.0500 +/- 0.56981",
      "basePath": "/home/steve/bench-results/qwen36-q4_0-gguf/correctness/fusemmvq2-ppl-single2-20260506T031402Z"
    },
    "tripleB70Generation": {
      "selector": "level_zero:2,1,3",
      "temperature": 0,
      "mode": "forced non-conversation completion",
      "controlSha256": "5f8395bd9f2bbccb6fa030b2fc173e52ed7035621d7cb20ebc24589fe9dbcbcf",
      "fusedSha256": "5f8395bd9f2bbccb6fa030b2fc173e52ed7035621d7cb20ebc24589fe9dbcbcf",
      "diffLines": 0,
      "basePath": "/home/steve/bench-results/qwen36-q4_0-gguf/correctness/fusemmvq2-nocnv-3x-20260506T030459Z"
    },
    "tripleB70PplCaution": {
      "fusedComparison": {
        "controlPpl": "2.1800 +/- 0.70030",
        "fusedPpl": "2.2064 +/- 0.69888",
        "basePath": "/home/steve/bench-results/qwen36-q4_0-gguf/correctness/fusemmvq2-ppl-3x-20260506T031125Z"
      },
      "controlRepeat": {
        "firstPpl": "2.1587 +/- 0.69054",
        "secondPpl": "2.3188 +/- 0.78117",
        "basePath": "/home/steve/bench-results/qwen36-q4_0-gguf/correctness/ppl-3x-off-repeat-20260506T031519Z"
      },
      "interpretation": "3-card PPL is not currently a stable correctness oracle because the fusion-disabled baseline varies materially across repeat runs."
    }
  },
  "artifacts": {
    "note": "/home/steve/llm-optimization-artifacts/notes/2026-05-06-q4-fused-mmvq2.md",
    "patch": "/home/steve/llm-optimization-artifacts/patches/llama-cpp-sycl-fuse-mmvq2-current-20260506.patch.gz.b64",
    "localmaxxingResponse": "/tmp/localmaxxing-fusemmvq2-submit.json"
  }
}
