{
  "date": "2026-05-06",
  "model": "unsloth/Qwen3.6-27B",
  "quantization": "Q4_0",
  "engine": "llama.cpp SYCL Level Zero",
  "codeChange": {
    "file": "/home/steve/src/llama.cpp-q4-b70/ggml/src/ggml-sycl/ggml-sycl.cpp",
    "env": "GGML_SYCL_COMM_SYNC_AFTER",
    "mode1": "wait_all_streams() after single-kernel allreduce",
    "mode2": "reduce.wait() after single-kernel allreduce"
  },
  "threeGpu": {
    "selector": "level_zero:2,1,3",
    "devices": "SYCL0/SYCL1/SYCL2",
    "tensorSplit": "1/1/1",
    "syncAfterMode": 2,
    "correctness": {
      "status": "pass",
      "steps": 16,
      "path": "/home/steve/bench-results/qwen36-q4_0-gguf/correctness/syncafter2-reducewait-20260506T055151Z"
    },
    "benchmark": {
      "promptTokens": 512,
      "outputTokens": 512,
      "repeats": 3,
      "promptTokS": 118.722431,
      "decodeTokS": 46.194319,
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-syncafter2-reducewait-fusemmvq2-triple213-p512n512-r3-20260506T055402Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-syncafter2-reducewait-fusemmvq2-triple213-p512n512-r3-20260506T055402Z.log"
    },
    "localmaxxing": {
      "id": "cmotnyi25001jqu01fccla8cf",
      "tokSOut": 46.194319,
      "tokSTotal": 66.4,
      "engineFlagsStatus": "POST with engineFlags.commandSnippet returned 500; accepted row omits engineFlags and carries flags in notes."
    }
  },
  "fourGpu": {
    "selector": "level_zero:2,1,3,0",
    "devices": "SYCL0/SYCL1/SYCL2/SYCL3",
    "tensorSplit": "1/1/1/1",
    "syncAfterMode": 2,
    "correctness": {
      "status": "pass",
      "steps": 8,
      "path": "/home/steve/bench-results/qwen36-q4_0-gguf/correctness/syncafter2-reducewait-quad-20260506T055606Z"
    },
    "benchmark": {
      "promptTokens": 512,
      "outputTokens": 512,
      "repeats": 3,
      "promptTokS": 86.547242,
      "decodeTokS": 34.929313,
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-syncafter2-reducewait-fusemmvq2-quad2130-p512n512-r3-20260506T055829Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-syncafter2-reducewait-fusemmvq2-quad2130-p512n512-r3-20260506T055829Z.log"
    },
    "localmaxxing": {
      "id": "cmotpxzii000pqy019v46swqn",
      "tokSOut": 34.929313,
      "tokSTotal": 49.748,
      "status": "submitted_as_scaling_datapoint",
      "engineFlagsStatus": "Submitted without backend or engineFlags because LocalMaxxing rejects sycl-level-zero backend and returns 500 when engineFlags is present. Response still reports quantization as null despite posting Q4_0."
    },
    "rootSweepP512N256R2": [
      {"selector": "level_zero:0,1,2,3", "decodeTokS": 34.280733},
      {"selector": "level_zero:1,0,2,3", "decodeTokS": 34.745079},
      {"selector": "level_zero:2,1,3,0", "decodeTokS": 34.797031},
      {"selector": "level_zero:3,2,1,0", "decodeTokS": 35.068186}
    ],
    "rootSweepSummary": "/home/steve/bench-results/qwen36-q4_0-gguf/root-sweep-syncafter2-quad-p512n256-r2-20260506T060109Z.jsonl"
  },
  "conclusion": "reduce.wait() is a small 3x improvement and preserves correctness, but 4x remains bottlenecked by the single-root collective design."
}