{
  "date": "2026-07-18",
  "status": "promoted",
  "model": "0xSero/DeepSeek-V4-Flash-180B",
  "model_revision": "7c360e1cd4a5168099dbc54d16d929bf6df04990",
  "mode": "single-active-generation TP4+EP DSpark7",
  "hardware": "4x Intel Arc Pro B70 32GB",
  "artifact": "/mnt/fast-ai/bench-results/deepseek-v4-flash-xpu/dspark7-sharded-target-argmax-candidate-20260718T2100Z",
  "vllm_commit": "264c7f2f7df21ddeeab32ecca0353133344f1ac9",
  "xpu_kernel_commit": "31315673737d95da0f79179c8f755260ef02c1d6",
  "oneccl_commit": "48fda4f0e074db005596d6899d5227d3f0316c12",
  "strict_suite_medians_tok_s": [
    80.82005189243556,
    76.90017809136465,
    78.28722593298039
  ],
  "strict_suite_median_of_medians_tok_s": 78.28722593298039,
  "record_tok_s": 80.82005189243556,
  "record_p10_tok_s": 71.66955635980304,
  "record_mean_tok_s": 80.0334096575548,
  "record_full_after_ttft_tok_s": 80.74548522063007,
  "record_wall_tok_s": 67.76281785757796,
  "record_ttft_ms": 342.77765700244345,
  "realistic_requests": 36,
  "realistic_requests_cached_zero": 36,
  "exact_canary_suites": 4,
  "exact_canary_requests": 24,
  "exact_canary_passes": 24,
  "mechanism": "Guarded greedy target verification projects only the local TP vocabulary shard, selects local top-1, gathers tiny value/index pairs, and commits target tokens through a native SYCL rejection op instead of materializing and all-gathering the full 129280-token logits tensor.",
  "single_active_generation": true,
  "aggregate_throughput": false,
  "target_model_verified_accepted_tokens": true,
  "prior_record_tok_s": 80.1635783863126,
  "absolute_gain_tok_s": 0.6564735061229665,
  "relative_gain_percent": 0.819000323254056,
  "localmaxxing_id": "cmrquta9905w3lg013m5vxoqx"
}