{
  "base_url": "http://127.0.0.1:39118/v1",
  "completed": 3,
  "concurrency": 1,
  "context_distribution": {
    "4096": 3
  },
  "context_seed": 0,
  "context_tokens": 4096,
  "engine": "sglang-glm53-sm120-v0.1.1-rc14",
  "failed": 0,
  "failures": [],
  "finished_at": "2026-09-02T10:02:32Z",
  "gpu": "2x RTX PRO 6000 Blackwell Max-Q TP2 WSL2",
  "manifest": null,
  "max_context_tokens": 131072,
  "max_tokens": 256,
  "measurement_protocol": "capacity-v3",
  "metrics": {
    "content_chunks": 126,
    "content_chunks_s": null,
    "decode_tok_s_p50": 74.09986124878638,
    "decode_tok_s_p95": 75.81721000951062,
    "e2e_p50_ms": 931.509800022468,
    "e2e_p95_ms": 1337.2750999988057,
    "effective_prefill_tok_s_p50": 12922.511500837718,
    "effective_prefill_tok_s_p95": 16539.77260828992,
    "generation_p50_ms": 527.5847000302747,
    "generation_p95_ms": 701.7556999926455,
    "mean_inter_token_latency_ms_p50": 13.495301922935491,
    "mean_inter_token_latency_ms_p95": 13.74960588188568,
    "output_token_sources": {
      "usage": 3
    },
    "output_tokens": 129,
    "prefix_cache_hit_avg": 0.9915796564499831,
    "prompt_token_samples": 3,
    "prompt_tokens": 8907,
    "prompt_tokens_p50": 2969.0,
    "prompt_tokens_p95": 2969.0,
    "reasoning_chunks": 0,
    "throughput_tok_s": 44.22523507732214,
    "time_to_first_output_p50_ms": 229.75410002982244,
    "time_to_first_output_p95_ms": 809.690399968531,
    "ttft_p50_ms": 229.75460003362969,
    "ttft_p95_ms": 809.6911999746226,
    "ttft_samples": 3
  },
  "model": "glm53-flash-ormandj-sglang-sm120-tp2-131k-nospec",
  "request_timings": [
    {
      "content_chunks": 40,
      "decode_tok_s": 75.81721000951062,
      "e2e_ms": 1337.2750999988057,
      "effective_prefill_tok_s": 3666.8336442119007,
      "generation_ms": 527.5847000302747,
      "mean_inter_token_latency_ms": 13.189617500756867,
      "output_token_source": "usage",
      "output_tokens": 41,
      "planned_context_tokens": 4096,
      "prompt_tokens": 2969,
      "reasoning_chunks": 0,
      "request_index": 0,
      "time_to_first_output_ms": 809.690399968531,
      "ttft_ms": 809.6911999746226,
      "visible_generation_ms": 527.5839000241831
    },
    {
      "content_chunks": 34,
      "decode_tok_s": 72.72935737870442,
      "e2e_ms": 646.9933000043966,
      "effective_prefill_tok_s": 16539.77260828992,
      "generation_ms": 467.4865999841131,
      "mean_inter_token_latency_ms": 13.74960588188568,
      "output_token_source": "usage",
      "output_tokens": 35,
      "planned_context_tokens": 4096,
      "prompt_tokens": 2969,
      "reasoning_chunks": 0,
      "request_index": 1,
      "time_to_first_output_ms": 179.50670002028346,
      "ttft_ms": 179.50710002332926,
      "visible_generation_ms": 467.48619998106733
    },
    {
      "content_chunks": 52,
      "decode_tok_s": 74.09986124878638,
      "e2e_ms": 931.509800022468,
      "effective_prefill_tok_s": 12922.511500837718,
      "generation_ms": 701.7556999926455,
      "mean_inter_token_latency_ms": 13.495301922935491,
      "output_token_source": "usage",
      "output_tokens": 53,
      "planned_context_tokens": 4096,
      "prompt_tokens": 2969,
      "reasoning_chunks": 0,
      "request_index": 2,
      "time_to_first_output_ms": 229.75410002982244,
      "ttft_ms": 229.75460003362969,
      "visible_generation_ms": 701.7551999888383
    }
  ],
  "requests": 3,
  "run_id": "benchmark-20260902T100229Z",
  "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": "2026-09-02T10:02:29Z",
  "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": 2916.8866999680176
}