{
  "date": "2026-05-05",
  "model": "Qwen/Qwen3.6-27B",
  "localModel": "/home/steve/models/qwen3.6-27b-q4_0-gguf/Qwen3.6-27B-Q4_0.gguf",
  "engine": "llama.cpp",
  "engineVersion": "db44417+sycl-fused-add",
  "quantization": "Q4_0",
  "hardware": {
    "gpu": "Intel Arc Pro B70 32GB",
    "gpuCount": 3,
    "selector": "level_zero:2,1,3",
    "cpu": "AMD EPYC 9015",
    "os": "Ubuntu 24.04.4 LTS, kernel 6.17.0-22-generic"
  },
  "env": {
    "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": "llama-bench -m Qwen3.6-27B-Q4_0.gguf -dev SYCL0/SYCL1/SYCL2 -sm tensor -ts 1/1/1 -p 512 -n 512 -r 3 -ngl 99 -fa 1 -ub 32 -ctk f16 -ctv f16 -t 8 --poll 50 -o jsonl",
  "result": {
    "promptTokS": 135.665459,
    "decodeTokS": 44.004344,
    "decodeStddevTokS": 0.012942,
    "decodeSamplesTokS": [
      43.9985,
      44.0192,
      43.9953
    ],
    "totalTokS": 66.453788,
    "promptTokens": 512,
    "outputTokens": 512,
    "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-fuseadd-validate-triple213-p512n512-r3-20260505T025706Z.jsonl",
    "log": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-fuseadd-validate-triple213-p512n512-r3-20260505T025706Z.log"
  },
  "baseline": {
    "name": "Q4_0 TP3 event barrier",
    "decodeTokS": 43.605046,
    "localmaxxingId": "cmortp5vn000el404dj3zqv0u"
  },
  "localmaxxing": {
    "id": "cmos1jmsv000iih04iifehc8d",
    "payload": "reduced",
    "note": "full engineFlags payload returned HTTP 500"
  },
  "patch": {
    "path": "/home/steve/llm-optimization-artifacts/patches/llama-cpp-sycl-meta-fused-allreduce-add-20260505.patch",
    "sha256": "c4322255e33462a799d9bca1a33da66db7cafa9e7e37dd55c415d85a7c38cbf0",
    "focusedPath": "/home/steve/llm-optimization-artifacts/patches/llama-cpp-sycl-meta-fused-allreduce-add-focused-20260505.patch",
    "focusedSha256": "16d197831a426fa461c035fe90db972ebe2a54d9ef8d6b14b319539551bfa427"
  },
  "quality": {
    "qualityPreserving": true,
    "notes": "Fuses allreduce(partial) and residual ADD without changing model weights, quantization, KV dtype, sampling, speculative decode, or power limits."
  }
}