{
  "date": "2026-05-06",
  "model": {
    "hfId": "unsloth/Qwen3.6-27B",
    "localPath": "/home/steve/models/qwen3.6-27b-q4_0-gguf/Qwen3.6-27B-Q4_0.gguf",
    "quantization": "Q4_0"
  },
  "hardware": {
    "gpuName": "Intel Arc Pro B70 32GB",
    "gpuCount": 3,
    "selector": "level_zero:2,1,3",
    "os": "Ubuntu 24.04.4 LTS",
    "kernel": "6.17.0-23-generic",
    "gucFirmware": "xe/bmg_guc_70.bin 70.49.4"
  },
  "engine": {
    "name": "llama.cpp",
    "backend": "sycl",
    "buildDir": "/home/steve/src/llama.cpp-q4-b70/build-sycl-2026-bmg-g31",
    "commitLabel": "db44417"
  },
  "quality": {
    "weights": "same Q4_0 GGUF",
    "kvCache": "f16/f16",
    "speculativeDecode": false,
    "powerLimitChange": false
  },
  "patch": {
    "envGate": "GGML_META_FUSE_ALLREDUCE_GET_ROWS=1",
    "description": "Experimental fused allreduce plus GET_ROWS helper for small f32 mirrored partials",
    "patchFile": "/home/steve/llm-optimization-artifacts/patches/llama-cpp-sycl-meta-getrows-fusion-current-20260506.patch",
    "patchFileGzipBase64": "/home/steve/llm-optimization-artifacts/patches/llama-cpp-sycl-meta-getrows-fusion-current-20260506.patch.gz.b64"
  },
  "runs": [
    {
      "name": "getrows-fusion-probe",
      "promptTokens": 0,
      "outputTokens": 1,
      "repetitions": 1,
      "resultTokSOut": 26.774031,
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-getrows-fuse-probe-triple213-p0n1-20260506T000738Z.log",
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-getrows-fuse-probe-triple213-p0n1-20260506T000738Z.jsonl",
      "observations": {
        "backendAddPaths": 254,
        "backendGetRowsPaths": 2,
        "plainBackendPaths": 0,
        "capturedNode": "attn_output-63 -> GET_ROWS"
      }
    },
    {
      "name": "getrows-fusion-short-screen",
      "promptTokens": 512,
      "outputTokens": 128,
      "repetitions": 2,
      "tokSPrompt": 135.504744,
      "tokSOut": 45.129276,
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-getrows-fuse-triple213-p512n128-r2-20260506T000850Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-getrows-fuse-triple213-p512n128-r2-20260506T000850Z.log"
    },
    {
      "name": "getrows-fusion-full-validation",
      "promptTokens": 512,
      "outputTokens": 512,
      "repetitions": 3,
      "tokSPrompt": 135.626208,
      "tokSOut": 45.375471,
      "tokSTotalComputed": 68.00050808306554,
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-getrows-fuse-triple213-p512n512-r3-20260506T001019Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-getrows-fuse-triple213-p512n512-r3-20260506T001019Z.log"
    },
    {
      "name": "same-build-gate-off-control",
      "promptTokens": 512,
      "outputTokens": 512,
      "repetitions": 3,
      "tokSPrompt": 135.671512,
      "tokSOut": 45.340867,
      "tokSTotalComputed": 67.96732925410481,
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-getrows-control-gateoff-triple213-p512n512-r3-20260506T001228Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-getrows-control-gateoff-triple213-p512n512-r3-20260506T001228Z.log"
    },
    {
      "name": "current-best-reference",
      "promptTokens": 512,
      "outputTokens": 512,
      "repetitions": 3,
      "tokSPrompt": 135.469357,
      "tokSOut": 45.624065,
      "tokSTotalComputed": 68.259384,
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-postreboot-reshapeadd-triple213-p512n512-r3-20260505T233641Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-postreboot-reshapeadd-triple213-p512n512-r3-20260505T233641Z.log",
      "localmaxxingId": "cmot9sgsi000lib042rqd6c62"
    },
    {
      "name": "current-stack-getrows-off-decode-only",
      "promptTokens": 0,
      "outputTokens": 512,
      "repetitions": 3,
      "tokSOut": 48.628094,
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/next-fusion-probes-20260506/getrows-off-tp3-rmsmul-p0n512-r3-20260506T213804Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/next-fusion-probes-20260506/getrows-off-tp3-rmsmul-p0n512-r3-20260506T213804Z.log"
    },
    {
      "name": "current-stack-getrows-on-decode-only",
      "promptTokens": 0,
      "outputTokens": 512,
      "repetitions": 3,
      "tokSOut": 49.043584,
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/next-fusion-probes-20260506/getrows-on-tp3-rmsmul-p0n512-r3-20260506T213933Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/next-fusion-probes-20260506/getrows-on-tp3-rmsmul-p0n512-r3-20260506T213933Z.log"
    },
    {
      "name": "current-stack-getrows-on-full-validation-r5",
      "promptTokens": 512,
      "outputTokens": 512,
      "repetitions": 5,
      "tokSPrompt": 197.252755,
      "tokSOut": 49.403656,
      "tokSTotalComputed": 79.016858,
      "stddevTokSOut": 0.361676,
      "samplesTokSOut": [49.98, 49.3727, 49.4262, 49.2388, 49.0006],
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/next-fusion-probes-20260506/getrows-on-tp3-rmsmul-p512n512-r5-20260506T214555Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/next-fusion-probes-20260506/getrows-on-tp3-rmsmul-p512n512-r5-20260506T214555Z.log",
      "localmaxxingId": "cmoultsa900h0ld011f0r2hcs",
      "localmaxxingResponse": "/home/steve/bench-results/localmaxxing-qwen36-q4-getrows-3x-min-20260506T220352Z.json"
    },
    {
      "name": "current-stack-getrows-off-full-control-r5",
      "promptTokens": 512,
      "outputTokens": 512,
      "repetitions": 5,
      "tokSPrompt": 196.860926,
      "tokSOut": 48.827917,
      "stddevTokSOut": 0.07273,
      "samplesTokSOut": [48.9259, 48.7428, 48.8062, 48.8759, 48.7888],
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/next-fusion-probes-20260506/getrows-off-tp3-rmsmul-p512n512-r5-20260506T214819Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/next-fusion-probes-20260506/getrows-off-tp3-rmsmul-p512n512-r5-20260506T214819Z.log"
    }
  ],
  "correctness": {
    "method": "llama-completion -no-cnv greedy decode, same prompt and seed, current fused stack with GET_ROWS off vs on",
    "baselineSha256": "2039492ece1be609e945c074396527ae6e0bcaddd2cf82cce6fd847355711214",
    "getRowsSha256": "2039492ece1be609e945c074396527ae6e0bcaddd2cf82cce6fd847355711214",
    "outputsIdentical": true,
    "outputText": "333333333333333",
    "artifactDir": "/home/steve/bench-results/qwen36-q4_0-gguf/correctness/getrows-20260506T220108Z-completion-nocnv2"
  },
  "conclusion": "GET_ROWS fusion was neutral on the early stack but is a small reproducible current-stack win. Keep GGML_META_FUSE_ALLREDUCE_GET_ROWS=1 in the TP3 best Q4_0 recipe, while continuing deeper work on projection GEMV plus allreduce/residual epilogues."
}