{
  "date": "2026-05-06",
  "model": {
    "hfIdSubmitted": "unsloth/Qwen3.6-27B",
    "localPath": "/home/steve/models/qwen3.6-27b-q4_0-gguf/Qwen3.6-27B-Q4_0.gguf",
    "quantization": "Q4_0",
    "kvCacheDtype": "f16"
  },
  "hardware": {
    "gpuName": "Intel Arc Pro B70",
    "vramGbPerGpu": 32,
    "cpu": "AMD EPYC 9015 8-Core Processor",
    "ramGb": 16,
    "os": "Ubuntu 24.04.4 LTS"
  },
  "engine": {
    "name": "llama.cpp",
    "backend": "SYCL Level Zero",
    "commit": "db44417",
    "build": "/home/steve/src/llama.cpp-q4-b70/build-sycl-2026-bmg-g31"
  },
  "patch": {
    "envGate": "GGML_SYCL_FUSE_RMS_NORM_MUL=1",
    "artifact": "patches/llama-cpp-sycl-rmsnormmul-current-20260506.patch.gz.b64",
    "filesTouched": [
      "ggml/src/ggml-sycl/norm.hpp",
      "ggml/src/ggml-sycl/norm.cpp",
      "ggml/src/ggml-sycl/ggml-sycl.cpp",
      "ggml/src/ggml-backend-meta.cpp",
      "ggml/src/ggml-alloc.c"
    ]
  },
  "correctness": {
    "method": "llama-completion greedy decode, same prompt and seed, fused MMVQ2 and fused SwiGLU enabled on both sides",
    "baselineSha256": "f7254271342a273042f88b21af7267f2fe5a06340ba68a9fc765746090a645aa",
    "rmsNormMulSha256": "f7254271342a273042f88b21af7267f2fe5a06340ba68a9fc765746090a645aa",
    "debugFusedCallsOneToken": 418,
    "outputsIdentical": true
  },
  "commonEnv": {
    "GGML_SYCL_DISABLE_DNN": "1",
    "GGML_SYCL_Q8_CACHE": "1",
    "GGML_SYCL_FUSE_MMVQ2": "1",
    "GGML_SYCL_FUSE_MMVQ2_SWIGLU": "1",
    "GGML_SYCL_FUSE_RMS_NORM_MUL": "1",
    "GGML_SYCL_COMM_ALLREDUCE": "1",
    "GGML_SYCL_COMM_SINGLE_KERNEL": "1",
    "GGML_SYCL_COMM_EVENT_BARRIER": "1",
    "GGML_SYCL_COMM_SYNC_AFTER": "2",
    "GGML_META_FUSE_ALLREDUCE_ADD": "1"
  },
  "results": [
    {
      "label": "single-sycl2-rmsnormmul",
      "gpuCount": 1,
      "devices": "SYCL2",
      "promptTokens": 512,
      "outputTokens": 512,
      "decodeTokS": 24.960284,
      "promptTokS": 525.603352,
      "totalTokS": 47.655433,
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/rms-norm-mul-20260506/single-sycl2-p512n512/rmsmul-20260506T204001Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/rms-norm-mul-20260506/single-sycl2-p512n512/rmsmul-20260506T204001Z.log"
    },
    {
      "label": "tp2-sycl21-rmsnormmul-ub128",
      "gpuCount": 2,
      "devices": "SYCL2/SYCL1",
      "splitMode": "tensor",
      "tensorSplit": "1/1",
      "ubatch": 128,
      "promptTokens": 512,
      "outputTokens": 512,
      "decodeTokS": 42.106013,
      "decodeStddevTokS": 0.011783,
      "promptTokS": 368.495701,
      "totalTokS": 75.570584,
      "localMaxxingId": "cmouju3dx00f3ld01rzmp9u76",
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/rms-norm-mul-20260506/tensor2-sycl21-p512n512/rmsmul-p512n512-r3-ub128-20260506T205122Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/rms-norm-mul-20260506/tensor2-sycl21-p512n512/rmsmul-p512n512-r3-ub128-20260506T205122Z.log",
      "localMaxxingResponse": "/home/steve/bench-results/localmaxxing-qwen36-q4-rmsnormmul-2x-min-20260506T210807Z.json"
    },
    {
      "label": "tp3-sycl213-rmsnormmul-ub128",
      "gpuCount": 3,
      "devices": "SYCL2/SYCL1/SYCL3",
      "splitMode": "tensor",
      "tensorSplit": "1/1/1",
      "ubatch": 128,
      "promptTokens": 512,
      "outputTokens": 512,
      "decodeTokS": 49.366188,
      "decodeStddevTokS": 0.486931,
      "promptTokS": 206.347938,
      "totalTokS": 79.667255,
      "localMaxxingId": "cmoujcois00esld01c5s6bwht",
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/rms-norm-mul-20260506/tensor3-sycl213-p512n512/rmsmul-p512n512-r3-ub128-20260506T204554Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/rms-norm-mul-20260506/tensor3-sycl213-p512n512/rmsmul-p512n512-r3-ub128-20260506T204554Z.log",
      "localMaxxingResponse": "/home/steve/bench-results/localmaxxing-qwen36-q4-rmsnormmul-3x-simple-20260506T205435Z.json"
    }
  ],
  "failedDiagnostics": [
    {
      "label": "tp4-assist005-rmsnormmul",
      "gpuCount": 4,
      "selector": "level_zero:2,1,3,0",
      "tensorSplit": "1/1/1/0.05",
      "outcome": "failed",
      "reason": "UR_RESULT_ERROR_OUT_OF_DEVICE_MEMORY during MUL_MAT",
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/rms-norm-mul-20260506/tensor4-assist005-sg2-p512n512/rmsmul-p512n512-r2-ub32-20260506T204412Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/rms-norm-mul-20260506/tensor4-assist005-sg2-p512n512/rmsmul-p512n512-r2-ub32-20260506T204412Z.log"
    }
  ]
}
