{
  "date": "2026-05-08",
  "localmaxxingId": "cmowt5ciy00d0o201f1mcrg3q",
  "status": "APPROVED",
  "model": {
    "hfId": "MiniMaxAI/MiniMax-M2.7",
    "localGguf": "MiniMax-M2.7-UD-IQ4_XS-00001-of-00004.gguf",
    "quantization": "UD-IQ4_XS"
  },
  "hardware": {
    "hwClass": "DISCRETE_GPU",
    "gpuName": "Intel Arc Pro B70",
    "gpuCount": 4,
    "vramGb": 32,
    "cpu": "AMD EPYC 9015 8-Core Processor",
    "ramGb": 16,
    "os": "Ubuntu 24.04.4 LTS"
  },
  "engine": {
    "engineName": "llama.cpp",
    "engineVersion": "ik_llama.cpp 9a26522 + local RPC/SYCL fast IQ4_XS mul_mat_id patches",
    "backendSubmitted": null,
    "backendNote": "LocalMaxxing currently rejects xpu/sycl/rpc backend values; submitted without backend and recorded Level Zero/RPC details in engineFlags.extraFlags."
  },
  "metrics": {
    "promptTokens": 0,
    "outputTokens": 64,
    "contextLength": 512,
    "tokSOut": 17.335655,
    "tokSTotal": 17.335655,
    "samplesTokSOut": [17.1807, 17.3988, 17.4275],
    "stddevTokSOut": 0.134954
  },
  "comparison": {
    "previousHighTokS": 16.404929,
    "improvementPct": 5.67,
    "qualityScope": "Fast path matched existing SYCL output in synthetic IQ4_XS MUL_MAT_ID A/B. SYCL also matched a manual dequantized IQ4_XS oracle closely; the CPU backend graph path diverged from that oracle in this synthetic case."
  },
  "probe": {
    "source": "prototypes/minimax_mmid_iq4xs_check.cpp",
    "log": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax_mmid_iq4xs_check-ref-onoff-20260508T111055Z.log",
    "fastOff": {
      "cpuRefMaxAbs": 4.50142,
      "cpuRefMaxRel": 576.365,
      "cpuRefNmse": 0.0288577,
      "syclRefMaxAbs": 0.0794775,
      "syclRefMaxRel": 27.9215,
      "syclRefNmse": 0.0000144346,
      "syclCpuMaxAbs": 4.48206,
      "syclCpuMaxRel": 822.229,
      "syclCpuNmse": 0.031076,
      "syclSum": -63.626713506877422,
      "syclSumSq": 220952.15214172262,
      "first": [-4.7533946, -1.53130412, -4.87711143, 1.29914367, -6.31177616, -9.72053337, 1.87014472, -1.95123363]
    },
    "fastOn": {
      "cpuRefMaxAbs": 4.50142,
      "cpuRefMaxRel": 576.365,
      "cpuRefNmse": 0.0288577,
      "syclRefMaxAbs": 0.0794775,
      "syclRefMaxRel": 27.9215,
      "syclRefNmse": 0.0000144346,
      "syclCpuMaxAbs": 4.48206,
      "syclCpuMaxRel": 822.229,
      "syclCpuNmse": 0.031076,
      "syclSum": -63.626713506877422,
      "syclSumSq": 220952.15214172262,
      "first": [-4.7533946, -1.53130412, -4.87711143, 1.29914367, -6.31177616, -9.72053337, 1.87014472, -1.95123363]
    }
  },
  "commandSnippet": "GGML_SYCL_FAST_MUL_MAT_ID_IQ4_XS=1 GGML_DISABLE_FUSED_RMS_NORM=1 llama-bench -m MiniMax-M2.7-UD-IQ4_XS-00001-of-00004.gguf -rpc 127.0.0.1:50100|0,127.0.0.1:50101|0,127.0.0.1:50102|0,127.0.0.1:50103|0 -p 0 -n 64 -r 3 -ngl 99 -sm layer -ts 1/1/1/1 -fa 0 -nkvo 0 -ub 32 -ctk f16 -ctv f16 -t 4 -w 0 -mmp 1 -rtr 1 -fmoe 1 -no-mmad 0 -muge 1 -sas 0 -o json",
  "files": {
    "sourceJson": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/rpc-layer-fast-mmid-r3-p0n64-20260508T102809Z.jsonl",
    "probeLog": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax_mmid_iq4xs_check-ref-onoff-20260508T111055Z.log",
    "payload": "/home/steve/bench-results/localmaxxing-minimax-m27-fast-mmid-20260508.payload.json",
    "response": "/home/steve/bench-results/localmaxxing-minimax-m27-fast-mmid-20260508.response.json"
  },
  "notes": "Enables the default-off GGML_SYCL_FAST_MUL_MAT_ID_IQ4_XS path for MiniMax expert-down MUL_MAT_ID. This is a software scheduling/kernel change only: same UD-IQ4_XS weights, F16 KV, fused MoE/up-gate/MMAD settings, and no GPU power-limit change."
}