{
  "date": "2026-05-10",
  "host": "4x Intel Arc Pro B70 32GB, Ubuntu 24.04",
  "script": "benchmarks/b70_minimax_qk_boundary_bench.py",
  "command": "python -m torch.distributed.run --standalone --nproc-per-node=4 /home/steve/llm-optimizations-publish/benchmarks/b70_minimax_qk_boundary_bench.py",
  "environment": {
    "CCL_P2P": "1",
    "B70_MINIMAX_BOUNDARY_TOKENS": "1,72,512",
    "B70_MINIMAX_BOUNDARY_DTYPE": "float16"
  },
  "log": "/mnt/fast-ai/bench-results/minimax-m2.7-autoround-boundary/b70-minimax-qk-boundary-tp4-20260510T150340Z.jsonl",
  "shape": {
    "qSizePerRank": 1536,
    "kvSizePerRank": 256,
    "hiddenSize": 3072,
    "worldSize": 4
  },
  "results": [
    {
      "tokens": 1,
      "tiny_allreduce_ms": 0.0169,
      "qk_var_ms": 0.00382,
      "qk_var_allreduce_ms": 0.0261,
      "qk_full_helper_ms": 0.0297,
      "torch_qk_reference_ms": 0.1746,
      "hidden_allreduce_ms": 0.0145
    },
    {
      "tokens": 72,
      "tiny_allreduce_ms": 0.0149,
      "qk_var_ms": 0.00374,
      "qk_var_allreduce_ms": 0.0233,
      "qk_full_helper_ms": 0.028,
      "torch_qk_reference_ms": 0.1761,
      "hidden_allreduce_ms": 0.0206
    },
    {
      "tokens": 512,
      "tiny_allreduce_ms": 0.0151,
      "qk_var_ms": 0.00555,
      "qk_var_allreduce_ms": 0.0238,
      "qk_full_helper_ms": 0.02845,
      "torch_qk_reference_ms": 0.1777,
      "hidden_allreduce_ms": 0.1188
    }
  ],
  "interpretation": "The isolated Q/K helper is much faster than the torch reference at layer-sized shapes, but full vLLM runs regressed because the helper broke existing compiled INT4 GEMM/RMS fusion and kept the allreduce wait boundary."
}