{
  "date": "2026-07-15",
  "record_tok_s": 34.06712071262497,
  "oneccl_runtime_record_identity": {
    "path": "/home/steve/.venvs/deepseek-v4-xpu/lib/libccl.so.1",
    "package": "oneccl==2021.17.2",
    "sha256": "ace144a390a53720b2743844decf127661c942b56f3b414900b9d8c11461acc3"
  },
  "ll_workgroup_geometry": {
    "source_commit": "c6aec66",
    "library_sha256": "77e3914ed9b22a88501f91172096a38fdfe5b15770f8525cb21b7966ee4578fb",
    "collectives": 87,
    "dtype": "bfloat16",
    "elements": 4096,
    "all_correct": true,
    "device_ms": {
      "threads_64_control_a": 6.106464,
      "threads_4": 5.684614,
      "threads_8": 5.526326,
      "threads_16": 5.562752,
      "threads_32": 5.529992,
      "threads_64_control_b": 5.665504
    },
    "decision": "closed: best 8-thread result saves 0.360 ms versus mean controls and only 0.139 ms versus reverse control, below the 0.50 ms integration gate"
  },
  "recursive_doubling": {
    "source_commit": "bef2321aa2d11b188ad2341ec837167a353e6f52",
    "library_sha256": "5f739536d2ce615a550b1c31575d68e84e12e4d198497482557243e4e3c96e4b",
    "collectives": 87,
    "dtype": "bfloat16",
    "elements": 4096,
    "all_correct": true,
    "max_rank_device_ms": {
      "ring_control_a": 5.57973,
      "recursive_doubling": 5.64486,
      "ring_control_b": 5.578872
    },
    "decision": "rejected: two XOR rounds are 0.0656 ms slower than the mean paired ring controls"
  },
  "expert_round_robin": {
    "vllm_commit": "62b8bed9ae3bea5aae9baf543491d029ab2adb9e",
    "run": "/mnt/fast-ai/bench-results/deepseek-v4-flash-xpu/expert-map-round-robin-20260715T0336Z",
    "placement": "rank 0 owns global experts 0,4,8,...,156; analogous interleaving on ranks 1-3",
    "cached_tokens": [0, 0, 0],
    "changed_input_expected": [1073, 437, 1073],
    "changed_input_actual": [1369, 361, 1369],
    "decision": "correctness-rejected before speed testing; packed MXFP4 post-load state retains a contiguous-expert ownership assumption"
  },
  "rank_skew_profile": {
    "run": "/mnt/fast-ai/bench-results/deepseek-v4-flash-xpu/current-graph-rank-skew-profile-20260715T031852Z",
    "collective_events_per_rank": 87,
    "decision": "inconclusive: profiler startup and serialization distort cross-device timestamps and collective duration, so raw multi-GPU timestamp deltas are not normal decode rank-arrival evidence"
  },
  "ring_readiness_markers": {
    "source_commit": "1edec4576f6424f645fcee8b18df6844c636423c",
    "library_sha256": "eb77b800754754c9f9ede04bff6ab0850a8e82a398c0c23baa2d38a10bbaef3f",
    "selector": "CCL_SYCL_ALLREDUCE_LL=ring_markers",
    "collectives": 87,
    "dtype": "bfloat16",
    "elements": 4096,
    "changed_epochs": 24,
    "all_correct": true,
    "max_rank_device_ms": {
      "ring_control_a": 5.165472,
      "ring_markers": 5.204264,
      "ring_control_b": 5.395494
    },
    "worst_marker_tax_ms_per_87": 0.038792,
    "worst_marker_tax_us_per_boundary": 0.445885,
    "decision": "passed: release-ordered per-wire epoch publication costs at most 0.446 us per boundary against the faster paired control, below the 1 us gate; proceed to the resident consumer microgate"
  }
}