{
  "date": "2026-05-06",
  "model": "vrfai/Qwen3.6-27B-FP8",
  "modelPath": "/home/steve/models/qwen3.6-27b-fp8-vrfai",
  "engine": "vllm",
  "engineVersion": "0.20.1+xpu",
  "quantization": "fp8/compressed-tensors",
  "libraryOrderingFix": {
    "issue": "XCCL segfaulted at dist.barrier() when oneAPI setvars libraries preceded the vLLM venv libraries.",
    "fix": "Prepend /home/steve/.venvs/vllm-xpu-managed/lib to LD_LIBRARY_PATH for torch/vLLM runs and avoid oneAPI setvars for vLLM.",
    "wrapper": "/home/steve/bench-vllm-qwen36-fp8.sh"
  },
  "xcclGate": {
    "selector": "level_zero:0,1",
    "env": {
      "LD_LIBRARY_PATH_PREFIX": "/home/steve/.venvs/vllm-xpu-managed/lib",
      "CCL_ATL_TRANSPORT": "ofi"
    },
    "status": "passed",
    "payloadGBpsAt256MiB": 41.75,
    "log": "/home/steve/bench-results/qwen36-fp8-vllm/xccl-standalone-2rank-venvlib-post-reboot-20260506T072239Z.log"
  },
  "pp2Tp2NoSpec": {
    "status": "stable_but_slow",
    "topology": {
      "tensorParallel": 2,
      "pipelineParallel": 2,
      "gpuCount": 4
    },
    "promptTokens": 512,
    "outputTokens": 512,
    "contextLength": 1024,
    "batchSize": 1,
    "latenciesSec": [
      18.831783488996734,
      18.411271026998293,
      18.654142821003916
    ],
    "avgLatencySec": 18.63239911233298,
    "tokSOut": 27.479016,
    "tokSTotal": 54.958033,
    "json": "/home/steve/bench-results/qwen36-fp8-vllm/vllm-qwen36-fp8-compressed-tensors-tp2-pp2-in512-out512-bs1-20260506T072309Z.json",
    "log": "/home/steve/bench-results/qwen36-fp8-vllm/vllm-qwen36-fp8-compressed-tensors-tp2-pp2-in512-out512-bs1-20260506T072309Z.log"
  },
  "tp4NgramBest": {
    "status": "new_validated_best",
    "topology": {
      "tensorParallel": 4,
      "pipelineParallel": 1,
      "gpuCount": 4
    },
    "speculative": {
      "method": "ngram",
      "numSpeculativeTokens": 4,
      "promptLookupMin": 2,
      "promptLookupMax": 4
    },
    "promptTokens": 512,
    "outputTokens": 512,
    "contextLength": 1024,
    "batchSize": 1,
    "latenciesSec": [
      10.693160828996042,
      11.773327080001764,
      8.512563062999106
    ],
    "avgLatencySec": 10.326350323998971,
    "tokSOut": 49.581893,
    "tokSTotal": 99.163787,
    "json": "/home/steve/bench-results/qwen36-fp8-vllm/vllm-qwen36-fp8-compressed-tensors-tp4-pp1-in512-out512-bs1-20260506T072633Z.json",
    "log": "/home/steve/bench-results/qwen36-fp8-vllm/vllm-qwen36-fp8-compressed-tensors-tp4-pp1-in512-out512-bs1-20260506T072633Z.log",
    "localmaxxingId": "cmotql1v60013qy01016jcs7r"
  },
  "conclusion": "Current fastest FP8 path is TP4 with n-gram speculative decoding. PP2 x TP2 is stable at 512/512 but too slow without speculative decode."
}