{
  "label": "llamacpp-qwen36-27b-q4_0-sycl-tp3-reshapeadd-p512-n512",
  "dateUtc": "2026-05-05T12:55:51Z",
  "localmaxxingId": "cmosmudwl0004k004hzz6l4u6",
  "model": {
    "hfId": "Qwen/Qwen3.6-27B",
    "file": "/home/steve/models/qwen3.6-27b-q4_0-gguf/Qwen3.6-27B-Q4_0.gguf",
    "quantization": "Q4_0",
    "params": 26895998464
  },
  "hardware": {
    "gpu": "Intel Arc Pro B70 32GB",
    "gpuCount": 3,
    "selector": "level_zero:2,1,3",
    "cpu": "AMD EPYC 9015",
    "ramGb": 16,
    "os": "Ubuntu 24.04.4 LTS",
    "kernel": "6.17.0-23-generic",
    "guc": "70.49.4"
  },
  "engine": {
    "name": "llama.cpp",
    "commit": "db44417",
    "backend": "SYCL Level Zero",
    "versionLabel": "db44417-sycl-q8-fuseadd-reshapeadd",
    "build": "/home/steve/src/llama.cpp-q4-b70/build-sycl-2026-bmg-g31/bin/llama-bench"
  },
  "env": {
    "ONEAPI_DEVICE_SELECTOR": "level_zero:2,1,3",
    "GGML_SYCL_DISABLE_DNN": "1",
    "GGML_SYCL_Q8_CACHE": "1",
    "GGML_SYCL_ASYNC_CPY_TENSOR": "1",
    "GGML_SYCL_COMM_ALLREDUCE": "1",
    "GGML_SYCL_COMM_SINGLE_KERNEL": "1",
    "GGML_SYCL_COMM_EVENT_BARRIER": "1",
    "GGML_META_FUSE_ALLREDUCE_ADD": "1"
  },
  "command": "/home/steve/src/llama.cpp-q4-b70/build-sycl-2026-bmg-g31/bin/llama-bench -m /home/steve/models/qwen3.6-27b-q4_0-gguf/Qwen3.6-27B-Q4_0.gguf -dev SYCL0/SYCL1/SYCL2 -ngl 99 -sm tensor -ts 1/1/1 -fa 1 -ub 32 -ctk f16 -ctv f16 -t 8 -p 512 -n 512 -r 3 --poll 50 -o jsonl",
  "metrics": {
    "promptTokens": 512,
    "outputTokens": 512,
    "promptTokS": 135.806175,
    "decodeTokS": 44.812806,
    "totalTokS": 67.388784,
    "decodeStddevTokS": 0.074867,
    "decodeSamplesTokS": [44.8839, 44.8199, 44.7346],
    "previousControlDecodeTokS": 44.238455,
    "improvementVsControlPercent": 1.2983
  },
  "trace": {
    "fusedReshapeAddSites": 48,
    "backendAddAllreduces": 127,
    "plainBackendAllreduces": 1,
    "probeLog": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-reshapeadd-probe-triple213-p0n1-20260505T125150Z.log",
    "probeJsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-reshapeadd-probe-triple213-p0n1-20260505T125150Z.jsonl"
  },
  "artifacts": {
    "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-reshapeadd-triple213-p512n512-r3-20260505T125442Z.jsonl",
    "log": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-reshapeadd-triple213-p512n512-r3-20260505T125442Z.log",
    "patch": "/home/steve/llm-optimization-artifacts/patches/llama-cpp-sycl-current-q4-reshapeadd-20260505.patch",
    "patchSha256": "000d6b31da654069ea040c0202d47ff64292116d126bdb29082dde0b4a01210f",
    "focusedPatch": "/home/steve/llm-optimization-artifacts/patches/llama-cpp-meta-reshapeadd-focused-20260505.patch",
    "focusedPatchSha256": "2860d6f8bd5a1f568002041f33999065e147d15c6a0f60d984f5c732025d623a"
  },
  "qualityNotes": "Quality-preserving graph scheduling change: the skipped RESHAPE is view-only, and the fused helper computes the same allreduce sum plus mirrored residual into the same ADD output. No model, quantization, KV cache, speculative decoding, sampling, or power-limit change."
}