{
  "date": "2026-05-04",
  "model": {
    "hfId": "Qwen/Qwen3.6-27B",
    "localFile": "/home/steve/models/qwen3.6-27b-q4_0-gguf/Qwen3.6-27B-Q4_0.gguf",
    "quantization": "Q4_0"
  },
  "hardware": {
    "gpuName": "Intel Arc Pro B70",
    "gpuCount": 3,
    "vramGbPerGpu": 32,
    "selector": "level_zero:2,1,3",
    "cpu": "AMD EPYC 9015 8-Core Processor",
    "os": "Ubuntu 24.04.4 LTS"
  },
  "engine": {
    "name": "llama.cpp",
    "version": "db44417 local SYCL",
    "binary": "/home/steve/src/llama.cpp-q4-b70/build-sycl-2026-bmg-g31/bin/llama-bench",
    "backend": "SYCL Level Zero",
    "splitMode": "tensor",
    "tensorSplit": "1/1/1",
    "gpuLayers": 99,
    "flashAttention": true,
    "kvCacheDtype": "f16"
  },
  "patch": {
    "env": "GGML_SYCL_COMM_EVENT_BARRIER=1",
    "description": "Use ext_oneapi_submit_barrier({reduce}) on non-root allreduce queues instead of a dependent single_task marker in the experimental single-kernel SYCL allreduce path.",
    "qualityImpact": "None expected. The patch only changes synchronization marker submission after the same allreduce kernel; weights, KV dtype, sampling, and math are unchanged."
  },
  "screen": {
    "shape": "512 prompt / 128 output",
    "eventBarrierOffTokS": 41.983073,
    "eventBarrierOnTokS": 43.330808,
    "offJsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-eventbarrier-triple213-p512n128-eb0-20260504T231527Z.jsonl",
    "onJsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-eventbarrier-triple213-p512n128-eb1-20260504T231630Z.jsonl"
  },
  "validation": {
    "shape": "512 prompt / 512 output",
    "repeats": 3,
    "tokSOut": 43.605046,
    "tokSPrompt": 135.697394,
    "tokSTotal": 65.999626,
    "samplesTokS": [43.9488, 43.4813, 43.385],
    "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-eventbarrier-triple213-validate-p512n512-eb1-20260504T231752Z.jsonl",
    "log": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-eventbarrier-triple213-validate-p512n512-eb1-20260504T231752Z.log"
  },
  "localmaxxing": {
    "submitted": true,
    "id": "cmortp5vn000el404dj3zqv0u",
    "status": "APPROVED",
    "note": "Reduced payload accepted. Full engineFlags payload hit HTTP 500, so command details are stored in localmaxxing_payloads.json and this artifact."
  }
}