{
  "base_url": "http://127.0.0.1:39063/v1",
  "completed": 3,
  "concurrency": 1,
  "context_distribution": {
    "32768": 3
  },
  "context_seed": 0,
  "context_tokens": 32768,
  "engine": "vllm-r16-b12x",
  "failed": 0,
  "failures": [],
  "finished_at": "2026-08-02T11:07:20Z",
  "gpu": "2x-rtx-pro-6000-blackwell-max-q-tp2",
  "manifest": null,
  "max_context_tokens": 1000000,
  "max_tokens": 1024,
  "measurement_protocol": "capacity-v4-reasoning",
  "metrics": {
    "content_chunks": 30,
    "content_chunks_s": null,
    "decode_tok_s_p50": 12.97821088175102,
    "decode_tok_s_p95": 13.40327035385703,
    "e2e_p50_ms": 13201.970899999651,
    "e2e_p95_ms": 28914.66109999965,
    "effective_prefill_tok_s_p50": 7195.350155029865,
    "effective_prefill_tok_s_p95": 7643.455401676511,
    "generation_p50_ms": 9953.484099998605,
    "generation_p95_ms": 25041.972499999247,
    "mean_inter_token_latency_ms_p50": 77.05222307692075,
    "mean_inter_token_latency_ms_p95": 84.3515601694797,
    "output_token_sources": {
      "usage": 3
    },
    "output_tokens": 525,
    "prefix_cache_hit_avg": 0.1387296426228573,
    "prompt_token_samples": 3,
    "prompt_tokens": 70122,
    "prompt_tokens_p50": 23374.0,
    "prompt_tokens_p95": 23374.0,
    "reasoning_chunks": 134,
    "throughput_tok_s": 10.279827799530409,
    "time_to_first_output_p50_ms": 3248.4868000010465,
    "time_to_first_output_p95_ms": 3872.6886000004015,
    "ttft_p50_ms": 10523.971900000106,
    "ttft_p95_ms": 27276.49569999994,
    "ttft_samples": 3
  },
  "model": "deepseek-v4-flash-0731-r16-b12x-dspark5-maxseq1-tp2-1m",
  "request_timings": [
    {
      "content_chunks": 7,
      "decode_tok_s": 12.97821088175102,
      "e2e_ms": 28914.66109999965,
      "effective_prefill_tok_s": 6035.600177095978,
      "generation_ms": 25041.972499999247,
      "mean_inter_token_latency_ms": 77.05222307692075,
      "output_token_source": "usage",
      "output_tokens": 326,
      "planned_context_tokens": 32768,
      "prompt_tokens": 23374,
      "reasoning_chunks": 93,
      "request_index": 0,
      "time_to_first_output_ms": 3872.6886000004015,
      "ttft_ms": 27276.49569999994,
      "visible_generation_ms": 1638.165399999707
    },
    {
      "content_chunks": 11,
      "decode_tok_s": 11.85514527521238,
      "e2e_ms": 13201.970899999651,
      "effective_prefill_tok_s": 7195.350155029865,
      "generation_ms": 9953.484099998605,
      "mean_inter_token_latency_ms": 84.3515601694797,
      "output_token_source": "usage",
      "output_tokens": 119,
      "planned_context_tokens": 32768,
      "prompt_tokens": 23374,
      "reasoning_chunks": 29,
      "request_index": 1,
      "time_to_first_output_ms": 3248.4868000010465,
      "ttft_ms": 10523.971900000106,
      "visible_generation_ms": 2677.998999999545
    },
    {
      "content_chunks": 12,
      "decode_tok_s": 13.40327035385703,
      "e2e_ms": 8952.124899999035,
      "effective_prefill_tok_s": 7643.455401676511,
      "generation_ms": 5894.083899998805,
      "mean_inter_token_latency_ms": 74.60865696201019,
      "output_token_source": "usage",
      "output_tokens": 80,
      "planned_context_tokens": 32768,
      "prompt_tokens": 23374,
      "reasoning_chunks": 12,
      "request_index": 2,
      "time_to_first_output_ms": 3058.0410000002303,
      "ttft_ms": 6202.172600000267,
      "visible_generation_ms": 2749.9522999987676
    }
  ],
  "requests": 3,
  "run_id": "benchmark-20260802T110629Z",
  "schema": "anvil-serving.benchmark/v1",
  "serve_flags": {
    "no_thinking": false,
    "reasoning_effort": "low",
    "shared_prefix_burst": false,
    "thinking_mode": "default"
  },
  "source_recipe": null,
  "started_at": "2026-08-02T11:06:29Z",
  "tier": null,
  "timing_methodology": {
    "clock": "client time.perf_counter",
    "decode": "usage completion tokens after the first token divided by client-observed generation time; completion tokens may include reasoning tokens",
    "effective_prefill": "usage.prompt_tokens divided by client-observed time to first output; includes queueing, scheduling, prefill, and first-token work",
    "generation": "client-observed E2E minus time to first output",
    "mean_inter_token_latency": "client-observed generation time divided by completion tokens after the first token; not a raw per-token timestamp trace",
    "time_to_first_output": "request start through the first non-empty streamed reasoning or content delta",
    "ttft": "request start through the first non-empty streamed content delta"
  },
  "wall_clock_ms": 51070.894399999816
}