{
  "date": "2026-05-04",
  "model": {
    "localPath": "/home/steve/models/qwen3.6-27b-fp8-vrfai",
    "hfId": "vrfai/Qwen3.6-27B-FP8",
    "quantization": "compressed-tensors fp8"
  },
  "hardware": {
    "hwClass": "DISCRETE_GPU",
    "gpuName": "Intel Arc Pro B70 32GB",
    "gpuCount": 4,
    "vramGb": 32,
    "cpu": "AMD EPYC 9015 8-Core Processor",
    "ramGb": 15.2,
    "os": "Ubuntu 24.04"
  },
  "engine": {
    "name": "vllm",
    "version": "0.20.1",
    "backend": "xpu-fa2",
    "tensorParallel": 4,
    "kvCacheDtype": "auto",
    "languageModelOnly": true,
    "gpuMemUtil": 0.9
  },
  "patches": [
    "patches/vllm-xpu-qwen35-gdn-spec-fallback-contiguous-state.patch",
    "patches/bench-vllm-qwen36-fp8-wrapper-quantization.patch",
    "patches/bench-vllm-qwen36-fp8-wrapper-speculative-config.patch"
  ],
  "speculativeConfig": {
    "method": "ngram",
    "num_speculative_tokens": 2,
    "prompt_lookup_max": 5,
    "prompt_lookup_min": 2
  },
  "runs": [
    {
      "name": "ngram2_512p_256o_screen",
      "promptTokens": 512,
      "outputTokens": 256,
      "warmupIters": 1,
      "measuredIters": 2,
      "avgLatencySec": 6.05994565400033,
      "computedOutputTokS": 42.244603271484394,
      "computedTotalTokS": 126.7338098144532,
      "json": "/home/steve/bench-results/qwen36-fp8-vllm/vllm-qwen36-fp8-compressed-tensors-tp4-in512-out256-bs1-20260504T220138Z.json",
      "log": "/home/steve/bench-results/qwen36-fp8-vllm/vllm-qwen36-fp8-compressed-tensors-tp4-in512-out256-bs1-20260504T220138Z.log",
      "localmaxxingSubmitted": false
    },
    {
      "name": "ngram2_512p_512o_validation",
      "promptTokens": 512,
      "outputTokens": 512,
      "warmupIters": 1,
      "measuredIters": 3,
      "latenciesSec": [
        11.795413790998282,
        12.598177527994267,
        11.756794460001402
      ],
      "avgLatencySec": 12.050128592997984,
      "computedOutputTokS": 42.48917312778802,
      "computedTotalTokS": 84.97834625557604,
      "json": "/home/steve/bench-results/qwen36-fp8-vllm/vllm-qwen36-fp8-compressed-tensors-tp4-in512-out512-bs1-20260504T220350Z.json",
      "log": "/home/steve/bench-results/qwen36-fp8-vllm/vllm-qwen36-fp8-compressed-tensors-tp4-in512-out512-bs1-20260504T220350Z.log",
      "localmaxxingSubmitted": true,
      "localmaxxingId": "cmorr43b30004jj04h4hhb6v1",
      "localmaxxingStatus": "APPROVED"
    }
  ],
  "baselines": {
    "tp4Fp8Fa2_512p_256o_tokSOut": 39.264,
    "tp4Fp8Fa2_512p_512o_tokSOut": 41.503
  },
  "interpretation": "N-gram speculative decode is a modest quality-preserving speed path for the patched static FP8 TP4 path. It does not use FP8 KV and does not raise power limits. Backend is vLLM XPU/FA2; LocalMaxxing backend field was omitted because the API backend enum has no xpu value."
}