{
  "base_url": "http://127.0.0.1:39118/v1",
  "completed": 3,
  "concurrency": 1,
  "context_distribution": {
    "262144": 3
  },
  "context_seed": 0,
  "context_tokens": 262144,
  "engine": "sglang-glm53-sm120-v0.1.1-rc14",
  "failed": 0,
  "failures": [],
  "finished_at": "2026-09-02T11:14:17Z",
  "gpu": "2x RTX PRO 6000 Blackwell Max-Q TP2 WSL2",
  "manifest": null,
  "max_context_tokens": 393216,
  "max_tokens": 256,
  "measurement_protocol": "capacity-v3",
  "metrics": {
    "content_chunks": 63,
    "content_chunks_s": null,
    "decode_tok_s_p50": 102.41880883554646,
    "decode_tok_s_p95": 105.21563169780745,
    "e2e_p50_ms": 39107.24919999484,
    "e2e_p95_ms": 39397.82940002624,
    "effective_prefill_tok_s_p50": 5578.783369064824,
    "effective_prefill_tok_s_p95": 5586.105240407523,
    "generation_p50_ms": 415.75149999698624,
    "generation_p95_ms": 646.2918000179343,
    "mean_inter_token_latency_ms_p50": 9.7638315790774,
    "mean_inter_token_latency_ms_p95": 11.878614285628178,
    "output_token_sources": {
      "usage": 3
    },
    "output_tokens": 144,
    "prefix_cache_hit_avg": null,
    "prompt_token_samples": 3,
    "prompt_tokens": 648302,
    "prompt_tokens_p50": 216101.0,
    "prompt_tokens_p95": 216101.0,
    "reasoning_chunks": 0,
    "throughput_tok_s": 1.2242969681044664,
    "time_to_first_output_p50_ms": 38736.2235999899,
    "time_to_first_output_p95_ms": 38751.537600008305,
    "ttft_p50_ms": 38736.223999992944,
    "ttft_p95_ms": 38751.538300013635,
    "ttft_samples": 3
  },
  "model": "glm53-flash-ormandj-sglang-sm120-tp2-393k-c1-adaptive-mtp",
  "request_timings": [
    {
      "content_chunks": 28,
      "decode_tok_s": 105.21563169780745,
      "e2e_ms": 39397.82940002624,
      "effective_prefill_tok_s": 5576.552915927487,
      "generation_ms": 646.2918000179343,
      "mean_inter_token_latency_ms": 9.504291176734327,
      "output_token_source": "usage",
      "output_tokens": 69,
      "planned_context_tokens": 262144,
      "prompt_tokens": 216100,
      "reasoning_chunks": 0,
      "request_index": 0,
      "time_to_first_output_ms": 38751.537600008305,
      "ttft_ms": 38751.538300013635,
      "visible_generation_ms": 646.2911000126041
    },
    {
      "content_chunks": 18,
      "decode_tok_s": 84.18490372314643,
      "e2e_ms": 39101.20239999378,
      "effective_prefill_tok_s": 5586.105240407523,
      "generation_ms": 415.75149999698624,
      "mean_inter_token_latency_ms": 11.878614285628178,
      "output_token_source": "usage",
      "output_tokens": 36,
      "planned_context_tokens": 262144,
      "prompt_tokens": 216101,
      "reasoning_chunks": 0,
      "request_index": 1,
      "time_to_first_output_ms": 38685.450899996795,
      "ttft_ms": 38685.45150000136,
      "visible_generation_ms": 415.75089999241754
    },
    {
      "content_chunks": 17,
      "decode_tok_s": 102.41880883554646,
      "e2e_ms": 39107.24919999484,
      "effective_prefill_tok_s": 5578.783369064824,
      "generation_ms": 371.0256000049412,
      "mean_inter_token_latency_ms": 9.7638315790774,
      "output_token_source": "usage",
      "output_tokens": 39,
      "planned_context_tokens": 262144,
      "prompt_tokens": 216101,
      "reasoning_chunks": 0,
      "request_index": 2,
      "time_to_first_output_ms": 38736.2235999899,
      "ttft_ms": 38736.223999992944,
      "visible_generation_ms": 371.0252000018954
    }
  ],
  "requests": 3,
  "run_id": "benchmark-20260902T111219Z",
  "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": "2026-09-02T11:12:19Z",
  "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": 117618.52210003417
}