{
  "base_url": "http://127.0.0.1:39038/v1",
  "completed": 3,
  "concurrency": 1,
  "context_distribution": {
    "65536": 3
  },
  "context_seed": 0,
  "context_tokens": 65536,
  "engine": "sglang",
  "failed": 0,
  "failures": [],
  "finished_at": "2026-08-21T14:45:26Z",
  "gpu": "rtx-5090",
  "manifest": null,
  "max_context_tokens": 131072,
  "max_tokens": 256,
  "measurement_protocol": "capacity-v3",
  "metrics": {
    "content_chunks": 157,
    "content_chunks_s": null,
    "decode_tok_s_p50": 69.74728544380527,
    "decode_tok_s_p95": 69.8286027815199,
    "e2e_p50_ms": 11308.540500001982,
    "e2e_p95_ms": 11405.887300032191,
    "effective_prefill_tok_s_p50": 5936.303209589728,
    "effective_prefill_tok_s_p95": 5943.634230172196,
    "generation_p50_ms": 687.3974000336602,
    "generation_p95_ms": 918.0671999929473,
    "mean_inter_token_latency_ms_p50": 14.337475553879308,
    "mean_inter_token_latency_ms_p95": 14.344799999889801,
    "output_token_sources": {
      "usage": 3
    },
    "output_tokens": 160,
    "prefix_cache_hit_avg": null,
    "prompt_token_samples": 3,
    "prompt_tokens": 185034,
    "prompt_tokens_p50": 61678.0,
    "prompt_tokens_p95": 61681.0,
    "reasoning_chunks": 0,
    "throughput_tok_s": 4.741920010584259,
    "time_to_first_output_p50_ms": 10390.473300009035,
    "time_to_first_output_p95_ms": 10718.48989999853,
    "ttft_p50_ms": 10390.473800012842,
    "ttft_p95_ms": 10718.490700004622,
    "ttft_samples": 3
  },
  "model": "qwen38-27b-radixark-nvfp4-sglang-rtx5090-128k-mm",
  "request_timings": [
    {
      "content_chunks": 48,
      "decode_tok_s": 69.8286027815199,
      "e2e_ms": 11405.887300032191,
      "effective_prefill_tok_s": 5754.0754878174075,
      "generation_ms": 687.3974000336602,
      "mean_inter_token_latency_ms": 14.320779167367922,
      "output_token_source": "usage",
      "output_tokens": 49,
      "planned_context_tokens": 65536,
      "prompt_tokens": 61675,
      "reasoning_chunks": 0,
      "request_index": 0,
      "time_to_first_output_ms": 10718.48989999853,
      "ttft_ms": 10718.490700004622,
      "visible_generation_ms": 687.3966000275686
    },
    {
      "content_chunks": 45,
      "decode_tok_s": 69.74728544380527,
      "e2e_ms": 11022.339100018144,
      "effective_prefill_tok_s": 5943.634230172196,
      "generation_ms": 645.1863999245688,
      "mean_inter_token_latency_ms": 14.337475553879308,
      "output_token_source": "usage",
      "output_tokens": 46,
      "planned_context_tokens": 65536,
      "prompt_tokens": 61678,
      "reasoning_chunks": 0,
      "request_index": 1,
      "time_to_first_output_ms": 10377.152700093575,
      "ttft_ms": 10377.153500099666,
      "visible_generation_ms": 645.1855999184772
    },
    {
      "content_chunks": 64,
      "decode_tok_s": 69.7116725229827,
      "e2e_ms": 11308.540500001982,
      "effective_prefill_tok_s": 5936.303209589728,
      "generation_ms": 918.0671999929473,
      "mean_inter_token_latency_ms": 14.344799999889801,
      "output_token_source": "usage",
      "output_tokens": 65,
      "planned_context_tokens": 65536,
      "prompt_tokens": 61681,
      "reasoning_chunks": 0,
      "request_index": 2,
      "time_to_first_output_ms": 10390.473300009035,
      "ttft_ms": 10390.473800012842,
      "visible_generation_ms": 918.06669998914
    }
  ],
  "requests": 3,
  "run_id": "benchmark-20260821T144452Z",
  "schema": "anvil-serving.benchmark/v1",
  "serve_flags": {
    "no_thinking": false,
    "reasoning_effort": null,
    "shared_prefix_burst": false,
    "thinking_mode": "disabled"
  },
  "source_recipe": null,
  "started_at": "2026-08-21T14:44:52Z",
  "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": 33741.606700001284
}