{
  "date_utc": "2026-05-05",
  "host": {
    "os": "Ubuntu 24.04.4 LTS",
    "kernel": "6.17.0-23-generic",
    "cpu": "AMD EPYC 9015",
    "ram_gb": 16,
    "gpu": "Intel Arc Pro B70 32GB",
    "gpu_count": 4,
    "guc": "xe/bmg_guc_70.bin 70.49.4"
  },
  "model": {
    "hf_id": "Qwen/Qwen3.6-27B",
    "local_file": "/home/steve/models/qwen3.6-27b-q4_0-gguf/Qwen3.6-27B-Q4_0.gguf",
    "quantization": "Q4_0"
  },
  "quality_policy": {
    "speculative_decode": false,
    "kv_cache_dtype": "f16",
    "power_limit_changes": false,
    "sampling_changes": false,
    "notes": "Quality-preserving graph scheduling only: fused allreduce+ADD and reshape-through-ADD path."
  },
  "runtime_recovery": {
    "sycl_lz_devices": 4,
    "peer_read_test": "peer kernel read ok across 4 devices",
    "avoid_recovery_method": "PCI function reset / FLR; caused xe GT1 PF self-configuration timeouts and a stuck sycl-ls before reboot"
  },
  "common_env": {
    "GGML_SYCL_DISABLE_DNN": "1",
    "GGML_SYCL_Q8_CACHE": "1",
    "GGML_SYCL_ASYNC_CPY_TENSOR": "1",
    "GGML_SYCL_COMM_ALLREDUCE": "1",
    "GGML_SYCL_COMM_SINGLE_KERNEL": "1",
    "GGML_SYCL_COMM_EVENT_BARRIER": "1",
    "GGML_META_FUSE_ALLREDUCE_ADD": "1"
  },
  "results": [
    {
      "label": "tp3_reshapeadd_postreboot_p512_n512_r3",
      "selector": "level_zero:2,1,3",
      "devices": "SYCL0/SYCL1/SYCL2",
      "tensor_split": "1/1/1",
      "prompt_tokens": 512,
      "output_tokens": 512,
      "repetitions": 3,
      "prompt_tok_s": 135.469357,
      "decode_tok_s": 45.624065,
      "decode_samples_tok_s": [45.5129, 45.4826, 45.8768],
      "computed_total_tok_s": 68.25938436655314,
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-postreboot-reshapeadd-triple213-p512n512-r3-20260505T233641Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-postreboot-reshapeadd-triple213-p512n512-r3-20260505T233641Z.log",
      "localmaxxing_id": "cmot9sgsi000lib042rqd6c62"
    },
    {
      "label": "tp4_reshapeadd_postreboot_p512_n512_r1",
      "selector": "level_zero:0,1,2,3",
      "devices": "SYCL0/SYCL1/SYCL2/SYCL3",
      "tensor_split": "1/1/1/1",
      "prompt_tokens": 512,
      "output_tokens": 512,
      "repetitions": 1,
      "prompt_tok_s": 102.210613,
      "decode_tok_s": 34.375523,
      "decode_samples_tok_s": [34.3755],
      "computed_total_tok_s": 51.448022192026855,
      "jsonl": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-postreboot-reshapeadd-quad0123-p512n512-r1-20260505T234410Z.jsonl",
      "log": "/home/steve/bench-results/qwen36-q4_0-gguf/sycl-k617023-postreboot-reshapeadd-quad0123-p512n512-r1-20260505T234410Z.log",
      "localmaxxing_id": "cmota1fpx0001l404wepbjtb7",
      "interpretation": "Valid negative-scaling diagnostic: 4x is stable from clean boot but slower than 3x for single-session Q4_0."
    }
  ],
  "next_steps": [
    "Use 45.624 tok/s 3x as the current Q4_0 control.",
    "Fuse or eliminate the final attn_output-63 -> GET_ROWS plain allreduce if mathematically safe.",
    "Avoid more 4x root/order sweeps until the 20 KiB reduction count or implementation changes.",
    "Do not use PCI FLR as a recovery path on this xe/B70 stack."
  ]
}