{ "base_url": "http://127.0.0.1:8001/v1", "clock_domain_id": null, "completed": 3, "concurrency": 1, "configuration_fingerprint": null, "context_distribution": { "120000": 3 }, "context_seed": 0, "context_tokens": 120000, "controlled_output_policy": "observe", "engine": "sglang-glm53-sm120-v0.1.1-rc14", "failed": 0, "failures": [], "finished_at": "2026-09-08T05:18:14Z", "finished_at_unix_ns": 1788844694042683771, "gpu": "2x RTX PRO 6000 Blackwell Max-Q TP2 Linux", "manifest": null, "max_context_tokens": 393216, "max_tokens": 256, "measurement_protocol": "capacity-v3", "metric_population": { "aggregate_throughput_eligible": true, "attempted_requests": 3, "excluded_requests": 0, "failure_policy": "failed or contract-invalid responses retain sanitized timing and metadata but are excluded from performance distributions", "p99_interpretation": "descriptive-nearest-rank-only", "p99_interpretation_by_metric": { "decode_tok_s": "descriptive-nearest-rank-only", "e2e": "descriptive-nearest-rank-only", "effective_prefill_tok_s": "descriptive-nearest-rank-only", "tpot": "descriptive-nearest-rank-only", "ttft": "descriptive-nearest-rank-only" }, "p99_minimum_recommended_samples": 100, "p99_samples_by_metric": { "decode_tok_s": 3, "e2e": 3, "effective_prefill_tok_s": 3, "tpot": 3, "ttft": 3 }, "performance_eligible_requests": 3, "successful_requests": 3 }, "metrics": { "completion_tokens_max": 50.0, "completion_tokens_mean": 43.333333333333336, "completion_tokens_mean_ci95_high": 52.75584066654589, "completion_tokens_mean_ci95_low": 33.91082600012078, "completion_tokens_min": 34.0, "completion_tokens_p25": 34.0, "completion_tokens_p50": 46.0, "completion_tokens_p75": 50.0, "completion_tokens_p90": 50.0, "completion_tokens_p95": 50.0, "completion_tokens_p99": 50.0, "completion_tokens_samples": 3, "completion_tokens_stddev": 8.32666399786453, "content_chunks": 60, "content_chunks_s": null, "controlled_output_exact_adherence": 0, "controlled_output_exact_adherence_rate": null, "controlled_output_samples": 0, "controlled_output_unobservable": 0, "decode_tok_s_max": 161.03063903058668, "decode_tok_s_mean": 133.46253900691633, "decode_tok_s_mean_ci95_high": 161.1490644028041, "decode_tok_s_mean_ci95_low": 105.77601361102856, "decode_tok_s_min": 114.3294021162164, "decode_tok_s_p25": 114.3294021162164, "decode_tok_s_p50": 125.02757587394592, "decode_tok_s_p75": 161.03063903058668, "decode_tok_s_p90": 161.03063903058668, "decode_tok_s_p95": 161.03063903058668, "decode_tok_s_p99": 161.03063903058668, "decode_tok_s_samples": 3, "decode_tok_s_stddev": 24.466565648328395, "e2e_max_ms": 16804.67718699947, "e2e_mean_ci95_high_ms": 17024.7803563455, "e2e_mean_ci95_low_ms": 15993.478186320945, "e2e_mean_ms": 16509.129271333222, "e2e_min_ms": 15984.34956300116, "e2e_p25_ms": 15984.34956300116, "e2e_p50_ms": 16738.361063999037, "e2e_p75_ms": 16804.67718699947, "e2e_p90_ms": 16804.67718699947, "e2e_p95_ms": 16804.67718699947, "e2e_p99_ms": 16804.67718699947, "e2e_samples": 3, "e2e_stddev_ms": 455.6805501118796, "effective_prefill_tok_s_max": 6101.365090143995, "effective_prefill_tok_s_mean": 5953.001825710278, "effective_prefill_tok_s_mean_ci95_high": 6098.398851062734, "effective_prefill_tok_s_mean_ci95_low": 5807.604800357823, "effective_prefill_tok_s_min": 5878.337452564196, "effective_prefill_tok_s_p25": 5878.337452564196, "effective_prefill_tok_s_p50": 5879.3029344226425, "effective_prefill_tok_s_p75": 6101.365090143995, "effective_prefill_tok_s_p90": 6101.365090143995, "effective_prefill_tok_s_p95": 6101.365090143995, "effective_prefill_tok_s_p99": 6101.365090143995, "effective_prefill_tok_s_samples": 3, "effective_prefill_tok_s_stddev": 128.48726284685418, "generation_max_ms": 428.5861650023435, "generation_mean_ci95_high_ms": 460.7949296888397, "generation_mean_ci95_low_ms": 201.49620964629904, "generation_mean_ms": 331.14556966756936, "generation_min_ms": 204.929945000913, "generation_p25_ms": 204.929945000913, "generation_p50_ms": 359.9205989994516, "generation_p75_ms": 428.5861650023435, "generation_p90_ms": 428.5861650023435, "generation_p95_ms": 428.5861650023435, "generation_p99_ms": 428.5861650023435, "generation_samples": 3, "generation_stddev_ms": 114.5710605743007, "mean_inter_token_latency_ms_max": 8.746656428619255, "mean_inter_token_latency_ms_mean": 7.6516300984338015, "mean_inter_token_latency_ms_mean_ci95_high": 9.126526329793903, "mean_inter_token_latency_ms_mean_ci95_low": 6.1767338670737, "mean_inter_token_latency_ms_min": 6.209998333361, "mean_inter_token_latency_ms_p25": 6.209998333361, "mean_inter_token_latency_ms_p50": 7.998235533321147, "mean_inter_token_latency_ms_p75": 8.746656428619255, "mean_inter_token_latency_ms_p90": 8.746656428619255, "mean_inter_token_latency_ms_p95": 8.746656428619255, "mean_inter_token_latency_ms_p99": 8.746656428619255, "mean_inter_token_latency_ms_samples": 3, "mean_inter_token_latency_ms_stddev": 1.303364902350794, "mean_time_per_output_token_max_ms": 8.746656428619255, "mean_time_per_output_token_mean_ci95_high_ms": 9.126526329793903, "mean_time_per_output_token_mean_ci95_low_ms": 6.1767338670737, "mean_time_per_output_token_mean_ms": 7.6516300984338015, "mean_time_per_output_token_min_ms": 6.209998333361, "mean_time_per_output_token_p25_ms": 6.209998333361, "mean_time_per_output_token_p50_ms": 7.998235533321147, "mean_time_per_output_token_p75_ms": 8.746656428619255, "mean_time_per_output_token_p90_ms": 8.746656428619255, "mean_time_per_output_token_p95_ms": 8.746656428619255, "mean_time_per_output_token_p99_ms": 8.746656428619255, "mean_time_per_output_token_samples": 3, "mean_time_per_output_token_stddev_ms": 1.303364902350794, "output_token_sources": { "usage": 3 }, "output_tokens": 130, "prefix_cache_hit_avg": 0.03057875275250322, "prompt_token_samples": 3, "prompt_tokens": 288834, "prompt_tokens_max": 96280.0, "prompt_tokens_mean": 96278.0, "prompt_tokens_mean_ci95_high": 96280.26321305522, "prompt_tokens_mean_ci95_low": 96275.73678694478, "prompt_tokens_min": 96276.0, "prompt_tokens_p25": 96276.0, "prompt_tokens_p50": 96278.0, "prompt_tokens_p75": 96280.0, "prompt_tokens_p90": 96280.0, "prompt_tokens_p95": 96280.0, "prompt_tokens_p99": 96280.0, "prompt_tokens_samples": 3, "prompt_tokens_stddev": 2.0, "reasoning_chunks": 0, "request_canary_passed": 0, "request_canary_samples": 0, "throughput_performance_eligible": true, "throughput_tok_s": 2.6245311985354314, "time_to_first_output_max_ms": 16378.440464999585, "time_to_first_output_mean_ci95_high_ms": 16568.578765723556, "time_to_first_output_mean_ci95_low_ms": 15787.38863760775, "time_to_first_output_mean_ms": 16177.983701665653, "time_to_first_output_min_ms": 15779.419618000247, "time_to_first_output_p25_ms": 15779.419618000247, "time_to_first_output_p50_ms": 16376.091021997127, "time_to_first_output_p75_ms": 16378.440464999585, "time_to_first_output_p90_ms": 16378.440464999585, "time_to_first_output_p95_ms": 16378.440464999585, "time_to_first_output_p99_ms": 16378.440464999585, "time_to_first_output_samples": 3, "time_to_first_output_stddev_ms": 345.1686204764836, "tpot_max_ms": 8.746656428619255, "tpot_mean_ci95_high_ms": 9.126526329793903, "tpot_mean_ci95_low_ms": 6.1767338670737, "tpot_mean_ms": 7.6516300984338015, "tpot_min_ms": 6.209998333361, "tpot_p25_ms": 6.209998333361, "tpot_p50_ms": 7.998235533321147, "tpot_p75_ms": 8.746656428619255, "tpot_p90_ms": 8.746656428619255, "tpot_p95_ms": 8.746656428619255, "tpot_p99_ms": 8.746656428619255, "tpot_samples": 3, "tpot_stddev_ms": 1.303364902350794, "ttft_max_ms": 16378.440944998147, "ttft_mean_ci95_high_ms": 16568.579041955018, "ttft_mean_ci95_low_ms": 15787.389501375754, "ttft_mean_ms": 16177.984271665386, "ttft_min_ms": 15779.420488001051, "ttft_p25_ms": 15779.420488001051, "ttft_p50_ms": 16376.091381996957, "ttft_p75_ms": 16378.440944998147, "ttft_p90_ms": 16378.440944998147, "ttft_p95_ms": 16378.440944998147, "ttft_p99_ms": 16378.440944998147, "ttft_samples": 3, "ttft_stddev_ms": 345.1683608736418 }, "model": "glm53-flash-ormandj-sglang-sm120-tp2-393k-c1-adaptive-mtp", "performance_eligible": true, "prompt_cache_mode": "shared", "request_canaries": false, "request_timings": [ { "content_chunks": 12, "controlled_output": null, "decode_tok_s": 161.03063903058668, "e2e_ms": 15984.34956300116, "effective_prefill_tok_s": 6101.365090143995, "generation_ms": 204.929945000913, "mean_inter_token_latency_ms": 6.209998333361, "mean_time_per_output_token_ms": 6.209998333361, "output_token_source": "usage", "output_tokens": 34, "planned_context_tokens": 120000, "prompt_tokens": 96276, "reasoning_chunks": 0, "request_canary": null, "request_index": 0, "time_to_first_output_ms": 15779.419618000247, "tpot_ms": 6.209998333361, "ttft_ms": 15779.420488001051, "visible_content_truncated": false, "visible_generation_ms": 204.92907500010915 }, { "content_chunks": 22, "controlled_output": null, "decode_tok_s": 125.02757587394592, "e2e_ms": 16738.361063999037, "effective_prefill_tok_s": 5878.337452564196, "generation_ms": 359.9205989994516, "mean_inter_token_latency_ms": 7.998235533321147, "mean_time_per_output_token_ms": 7.998235533321147, "output_token_source": "usage", "output_tokens": 46, "planned_context_tokens": 120000, "prompt_tokens": 96278, "reasoning_chunks": 0, "request_canary": null, "request_index": 1, "time_to_first_output_ms": 16378.440464999585, "tpot_ms": 7.998235533321147, "ttft_ms": 16378.440944998147, "visible_content_truncated": false, "visible_generation_ms": 359.9201190008898 }, { "content_chunks": 26, "controlled_output": null, "decode_tok_s": 114.3294021162164, "e2e_ms": 16804.67718699947, "effective_prefill_tok_s": 5879.3029344226425, "generation_ms": 428.5861650023435, "mean_inter_token_latency_ms": 8.746656428619255, "mean_time_per_output_token_ms": 8.746656428619255, "output_token_source": "usage", "output_tokens": 50, "planned_context_tokens": 120000, "prompt_tokens": 96280, "reasoning_chunks": 0, "request_canary": null, "request_index": 2, "time_to_first_output_ms": 16376.091021997127, "tpot_ms": 8.746656428619255, "ttft_ms": 16376.091381996957, "visible_content_truncated": false, "visible_generation_ms": 428.58580500251264 } ], "requests": 3, "response_words": 0, "responses_received": 3, "run_id": "benchmark-20260908T051724Z", "schema": "anvil-serving.benchmark/v1", "serve_flags": { "no_thinking": false, "reasoning_effort": null, "shared_prefix_burst": false, "thinking_mode": "disabled" }, "shared_prefix_tokens": 8000, "source_recipe": null, "started_at": "2026-09-08T05:17:24Z", "started_at_unix_ns": 1788844644510023399, "tier": null, "timestamp_precision": "nanosecond-unix-fields; ISO labels are second precision", "timing_methodology": { "artifact_timestamps": "time.time_ns Unix timestamps; independent artifacts may align them only when the operator supplied the same explicit clock_domain_id", "clock": "client time.perf_counter", "controlled_output": "--response-words adherence is recorded; observe preserves historical behavior while strict rejects non-exact or truncated completions; visible response capture is capped at 8192 characters", "decode": "usage completion tokens after the first token divided by client-observed generation time; completion tokens may include reasoning tokens", "distributions": "request-level mean, sample standard deviation, normal-approximation 95% confidence interval for the mean, min, p25, p50, p75, p90, p95, p99, and max using the nearest-rank percentile method", "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": "legacy key for a TPOT proxy: client-observed generation time divided by completion tokens after the first token; not raw per-token ITL", "p99": "numeric nearest-rank p99 is always retained for compatibility; it is descriptive only below 100 eligible samples", "request_canaries": "when enabled, each response must begin with its own synthetic marker and contain no other ANVIL_REQ marker; bounded/truncated captures fail closed and are excluded from performance metrics", "time_to_first_output": "request start through the first non-empty streamed reasoning or content delta", "tpot": "client-observed mean time per output token after the first output; identical to the legacy mean_inter_token_latency key", "ttft": "request start through the first non-empty streamed content delta" }, "wall_clock_ms": 49532.65561199805 }