{
  "hfId": "Lasimeri/MiniMax-M2.7-int4-AutoRound",
  "modelRevision": "main",
  "hardware": {
    "hwClass": "DISCRETE_GPU",
    "gpuName": "Intel Arc Pro B70",
    "gpuCount": 4,
    "vramGb": 32,
    "cpu": "AMD EPYC 9015",
    "ramGb": 16,
    "os": "Ubuntu 24.04.4"
  },
  "engineName": "vllm",
  "engineVersion": "0.20.1-local",
  "quantization": "INT4 AutoRound W4A16",
  "promptTokens": 512,
  "outputTokens": 256,
  "contextLength": 2048,
  "batchSize": 1,
  "tokSOut": 34.578045,
  "tokSTotal": 103.734136,
  "peakVramGb": 28.11,
  "notes": "MiniMax M2.7 AutoRound W4A16 on 4x Intel Arc Pro B70 via vLLM/XPU TP4. Experimental unsigned llm-scaler u4 decode-only ESIMD INT4 MoE path with FP32 router weights. This run leaves CCL_ZE_IPC_EXCHANGE unset so oneCCL uses its default IPC exchange instead of the earlier forced pidfd path, giving a small improvement. No speculative decode, no expert dropping, no sampling change, no power-limit change.",
  "engineFlags": {
    "commandSnippet": "VLLM_XPU_USE_LLM_SCALER_MOE=1 CCL_IPC=default vllm bench throughput --backend vllm --model /mnt/corsair-external/llm-models/minimax-m2.7-int4-autoround --tokenizer /mnt/corsair-external/llm-models/minimax-m2.7-int4-autoround --trust-remote-code --dtype float16 --tensor-parallel-size 4 --distributed-executor-backend mp --max-model-len 2048 --max-num-batched-tokens 1024 --max-num-seqs 1 --dataset-name random --random-input-len 512 --random-output-len 256 --random-range-ratio 0 --num-prompts 1 --disable-log-stats",
    "tensorParallel": 4,
    "gpuLayers": 999,
    "contextLength": 2048,
    "concurrency": 1,
    "flashAttn": true,
    "chunkedPrefill": true,
    "prefillChunkSize": 1024,
    "contBatching": true,
    "attentionBackend": "flash_attn",
    "engineQuant": "inc",
    "specDecoding": false,
    "extraFlags": "XPU/Level Zero; CCL_ATL_TRANSPORT=ofi; CCL_TOPO_P2P_ACCESS=1; CCL_ZE_IPC_EXCHANGE unset/default; VLLM_XPU_ENABLE_XPU_GRAPH=0; VLLM_XPU_USE_LLM_SCALER_MOE=1; PYTHONPATH includes /home/steve/src/llm-scaler/vllm/custom-esimd-kernels-vllm/python; dtype=float16; llm-scaler u4 decode-only patch with FP32 route-weight kernel variant"
  }
}