{
  "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": "2026-08-21T14:49:57Z",
  "gpu": "rtx-5090",
  "manifest": null,
  "max_context_tokens": 70000,
  "max_tokens": 256,
  "measurement_protocol": "capacity-v3",
  "metrics": {
    "content_chunks": 58,
    "content_chunks_s": null,
    "decode_tok_s_p50": 117.1020210347157,
    "decode_tok_s_p95": 132.08635428306962,
    "e2e_p50_ms": 11528.385400073603,
    "e2e_p95_ms": 11635.743999970146,
    "effective_prefill_tok_s_p50": 5581.271345007558,
    "effective_prefill_tok_s_p95": 5590.006781190826,
    "generation_p50_ms": 452.59680005256087,
    "generation_p95_ms": 476.9607000052929,
    "mean_inter_token_latency_ms_p50": 8.539562265142658,
    "mean_inter_token_latency_ms_p95": 9.514735556311077,
    "output_token_sources": {
      "usage": 3
    },
    "output_tokens": 164,
    "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": 4.735668443708395,
    "time_to_first_output_p50_ms": 11051.42470006831,
    "time_to_first_output_p95_ms": 11183.147199917585,
    "ttft_p50_ms": 11051.42540007364,
    "ttft_p95_ms": 11183.147999923676,
    "ttft_samples": 3
  },
  "model": "qwen38-27b-radixark-nvfp4-sglang-rtx5090-128k-mm",
  "request_timings": [
    {
      "content_chunks": 19,
      "decode_tok_s": 117.1020210347157,
      "e2e_ms": 11635.743999970146,
      "effective_prefill_tok_s": 5514.994920254159,
      "generation_ms": 452.59680005256087,
      "mean_inter_token_latency_ms": 8.539562265142658,
      "output_token_source": "usage",
      "output_tokens": 54,
      "planned_context_tokens": 65536,
      "prompt_tokens": 61675,
      "reasoning_chunks": 0,
      "request_index": 0,
      "time_to_first_output_ms": 11183.147199917585,
      "ttft_ms": 11183.147999923676,
      "visible_generation_ms": 452.59600004646927
    },
    {
      "content_chunks": 18,
      "decode_tok_s": 105.10013589780799,
      "e2e_ms": 11461.781200021505,
      "effective_prefill_tok_s": 5590.006781190826,
      "generation_ms": 428.1631000339985,
      "mean_inter_token_latency_ms": 9.514735556311077,
      "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": 11033.618099987507,
      "ttft_ms": 11033.618799992837,
      "visible_generation_ms": 428.16240002866834
    },
    {
      "content_chunks": 21,
      "decode_tok_s": 132.08635428306962,
      "e2e_ms": 11528.385400073603,
      "effective_prefill_tok_s": 5581.271345007558,
      "generation_ms": 476.9607000052929,
      "mean_inter_token_latency_ms": 7.570804761988776,
      "output_token_source": "usage",
      "output_tokens": 64,
      "planned_context_tokens": 65536,
      "prompt_tokens": 61681,
      "reasoning_chunks": 0,
      "request_index": 2,
      "time_to_first_output_ms": 11051.42470006831,
      "ttft_ms": 11051.42540007364,
      "visible_generation_ms": 476.95999999996275
    }
  ],
  "requests": 3,
  "run_id": "benchmark-20260821T144922Z",
  "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": "2026-08-21T14:49:22Z",
  "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": 34630.80279994756
}