{
  "date": "2026-05-08",
  "localmaxxingId": "cmox103ol0040ml019yzs6gvs",
  "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",
    "ramGb": 16,
    "os": "Ubuntu 24.04.4 LTS"
  },
  "engine": {
    "engineName": "llama.cpp",
    "engineVersion": "ik_llama.cpp patched 2026-05-08",
    "backendSubmitted": null,
    "backendNote": "LocalMaxxing currently rejects xpu/sycl/rpc backend values; submitted without backend and recorded SYCL/RPC details in notes and engineFlags.extraFlags."
  },
  "metrics": {
    "promptTokens": 0,
    "outputTokens": 64,
    "contextLength": 2048,
    "tokSOut": 17.697772,
    "tokSTotal": 17.697772,
    "samplesTokSOut": [17.5369, 17.7524, 17.7386, 17.7282, 17.7328],
    "stddevTokSOut": 0.090407
  },
  "previousMmvY2Result": {
    "localmaxxingId": "cmowx1t6z000mml01v111mzvl",
    "tokSOut": 17.547020,
    "samplesTokSOut": [17.3265, 17.6006, 17.6046, 17.6047, 17.5987],
    "sourceJson": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax-fast-mmid-mmv-runtime2-r5-p0n64-20260508T124605Z.jsonl"
  },
  "fusedRmsRecheck": {
    "enabled": true,
    "sourceJson": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax-fast-mmid-mmv-runtime2-fusedrms-ub64-r5-p0n64-20260508T143014Z.jsonl",
    "correctnessSmokeDir": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/correctness/fused-rms-y2-smoke-20260508T143721Z",
    "correctnessSha256": "78e517594d13b4c3405739fda666369c2cfe6afaa111e1e75a0b5eb0a848d6fd",
    "recommendation": "Keep fused RMSNorm enabled in the current MiniMax GGUF recipe."
  },
  "contextRun": {
    "localmaxxingId": "cmox1gcxl0049ml01kiijqbpo",
    "promptTokens": 512,
    "outputTokens": 128,
    "promptTokS": 54.506141,
    "promptSamplesTokS": [44.71, 59.4252, 59.3832],
    "outputTokS": 17.693021,
    "outputSamplesTokS": [17.5465, 17.7482, 17.7844],
    "tokSTotal": 38.48946204904967,
    "sourceJson": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax-fast-mmid-mmv-runtime2-fusedrms-ub64-r3-p512n128-20260508T144635Z.jsonl",
    "payload": "/home/steve/bench-results/localmaxxing-minimax-m27-fusedrms-mmv-y2-ub64-p512n128-20260508.payload.json",
    "response": "/home/steve/bench-results/localmaxxing-minimax-m27-fusedrms-mmv-y2-ub64-p512n128-20260508.response.json",
    "supersedes": {
      "localmaxxingId": "cmowyq5tu001jml01b470i75g",
      "promptTokS": 50.905433,
      "outputTokS": 17.515510,
      "tokSTotal": 36.85431292616281,
      "sourceJson": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax-fast-mmid-mmv-runtime2-r3-p512n128-20260508T133014Z.jsonl"
    }
  },
  "comparison": {
    "sameBuildControlTokS": 17.198973,
    "sameBuildImprovementPct": 2.02,
    "previousSubmittedHighTokS": 17.547020,
    "previousSubmittedImprovementPct": 0.86,
    "qualityScope": "Scheduling/kernel changes only. Same UD-IQ4_XS weights, F16 KV cache, sampler path, split mode, and no GPU power-limit changes. Deterministic 16-token greedy generation smokes matched the earlier baselines byte-for-byte."
  },
  "timingDeltaMs": {
    "fastMmidBaseline": {
      "MUL_MAT": 115.366,
      "MOE_FUSED_UP_GATE": 95.922,
      "MUL_MAT_ID": 19.653
    },
    "mmvY2": {
      "MUL_MAT": 111.580,
      "MOE_FUSED_UP_GATE": 88.382,
      "MUL_MAT_ID": 18.798
    }
  },
  "negativeFollowups": [
    {
      "variant": "GGML_SYCL_MMV_Y_RUNTIME=8",
      "tokSOut": 17.238444,
      "samplesTokSOut": [16.6394, 17.5203, 17.5557],
      "sourceJson": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax-fast-mmid-mmv-runtime8-r3-p0n64-20260508T131435Z.jsonl"
    },
    {
      "variant": "GGML_SYCL_MMV_Y_RUNTIME=2 plus GGML_SYCL_MOE_IQ4_XS_MMV_Y=4",
      "tokSOut": 17.232041,
      "samplesTokSOut": [16.6258, 17.5346, 17.5357],
      "sourceJson": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax-fast-mmid-mmv2-moe4-r3-p0n64-20260508T132138Z.jsonl"
    }
  ],
  "microbatchSweep": [
    {
      "ubatch": 64,
      "repeats": 5,
      "tokSOut": 17.559741,
      "samplesTokSOut": [17.3921, 17.607, 17.6027, 17.5936, 17.6033],
      "sourceJson": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax-fast-mmid-mmv-runtime2-ub64-r5-p0n64-20260508T135521Z.jsonl",
      "localmaxxingSubmitted": false,
      "notes": "Tiny local best over -ub 32; not submitted because the delta is inside normal run-to-run noise."
    },
    {
      "ubatch": 128,
      "repeats": 3,
      "tokSOut": 17.502587,
      "samplesTokSOut": [17.3805, 17.6, 17.5272],
      "sourceJson": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax-fast-mmid-mmv-runtime2-ub128-r3-p0n64-20260508T134833Z.jsonl",
      "localmaxxingSubmitted": false
    }
  ],
  "flashAttentionCheck": {
    "variant": "-fa 1 -ub 64 p512/n128/r3",
    "status": "failed",
    "error": "SYCL RPC worker aborted with unsupported FLASH_ATTN_EXT: ggml_backend_sycl_graph_compute: error: op not supported fa-0 (FLASH_ATTN_EXT)",
    "sourceLog": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax-fast-mmid-mmv-runtime2-ub64-fa1-r3-p512n128-20260508T140323Z.log",
    "recommendation": "Keep MiniMax GGUF on -fa 0 until FLASH_ATTN_EXT is implemented for this SYCL/RPC path or the backend can fallback cleanly."
  },
  "dnnCheck": {
    "variant": "RPC workers without GGML_SYCL_DISABLE_DNN=1, Y=2, -ub 64, p0/n64/r3",
    "tokSOut": 17.213021,
    "samplesTokSOut": [16.578, 17.5222, 17.5389],
    "sourceJson": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax-fast-mmid-mmv-runtime2-dnnon-ub64-r3-p0n64-20260508T141516Z.jsonl",
    "recommendation": "Keep GGML_SYCL_DISABLE_DNN=1 in the current MiniMax GGUF recipe."
  },
  "schedulerSweep": [
    {
      "variant": "-sas 1 -t 4",
      "tokSOut": 17.718033,
      "samplesTokSOut": [17.5327, 17.7945, 17.7565, 17.7508, 17.7557],
      "sourceJson": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax-fusedrms-mmv-y2-ub64-sas1-r5-p0n64-20260508T151206Z.jsonl",
      "localmaxxingSubmitted": false,
      "notes": "Tiny local best; not submitted because the delta over the submitted fused-RMS row is inside noise."
    },
    {
      "variant": "-sas 1 -t 8",
      "tokSOut": 17.691594,
      "samplesTokSOut": [17.564, 17.7566, 17.7542],
      "sourceJson": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax-fusedrms-mmv-y2-ub64-sas1-t8-r3-p0n64-20260508T151907Z.jsonl",
      "localmaxxingSubmitted": false
    },
    {
      "variant": "-rtr 0 -sas 0 -t 4",
      "tokSOut": 17.659099,
      "samplesTokSOut": [17.4545, 17.7551, 17.7676],
      "sourceJson": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax-fusedrms-mmv-y2-ub64-rtr0-r3-p0n64-20260508T145807Z.jsonl",
      "localmaxxingSubmitted": false
    }
  ],
  "commandSnippet": "GGML_SYCL_FAST_MUL_MAT_ID_IQ4_XS=1 GGML_SYCL_MMV_Y_RUNTIME=2 GGML_SYCL_DISABLE_DNN=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 5 -ngl 99 -sm layer -ts 1/1/1/1 -fa 0 -nkvo 0 -ub 64 -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/minimax-fast-mmid-mmv-runtime2-fusedrms-ub64-r5-p0n64-20260508T143014Z.jsonl",
    "sameBuildControlJson": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax-fast-mmid-control-r5-p0n64-20260508T120407Z.jsonl",
    "compileTimeMmv2Json": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/minimax-fast-mmid-mmv2-r5-p0n64-20260508T115654Z.jsonl",
    "correctnessSmokeDir": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/correctness/mmv-y2-smoke-20260508T125856Z",
    "payload": "/home/steve/bench-results/localmaxxing-minimax-m27-fusedrms-mmv-y2-ub64-20260508.payload.json",
    "response": "/home/steve/bench-results/localmaxxing-minimax-m27-fusedrms-mmv-y2-ub64-20260508.response.json"
  },
  "notes": "Current MiniMax GGUF recipe: fast IQ4_XS MUL_MAT_ID, runtime MMV row grouping Y=2, DNN disabled, fused RMSNorm enabled, ubatch 64. Runtime MMV values are constrained to 1, 2, 4, or 8. Runtime value 2 reproduced the compile-time MMV_Y=2 direction; fused RMSNorm became positive after the row-packing/ubatch changes."
}