{ "base_url": "http://127.0.0.1:8001/v1", "clock_domain_id": null, "completed": 3, "concurrency": 1, "configuration_fingerprint": null, "context_distribution": { "262144": 3 }, "context_seed": 0, "context_tokens": 262144, "controlled_output_policy": "observe", "engine": "sglang-glm53-sm120-v0.1.1-rc14", "failed": 0, "failures": [], "finished_at": "2026-09-08T05:20:35Z", "finished_at_unix_ns": 1788844835359286848, "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": 57.0, "completion_tokens_mean": 48.666666666666664, "completion_tokens_mean_ci95_high": 57.743063406440534, "completion_tokens_mean_ci95_low": 39.590269926892795, "completion_tokens_min": 41.0, "completion_tokens_p25": 41.0, "completion_tokens_p50": 48.0, "completion_tokens_p75": 57.0, "completion_tokens_p90": 57.0, "completion_tokens_p95": 57.0, "completion_tokens_p99": 57.0, "completion_tokens_samples": 3, "completion_tokens_stddev": 8.020806277010642, "content_chunks": 65, "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": 148.65052912929124, "decode_tok_s_mean": 130.0224527946473, "decode_tok_s_mean_ci95_high": 148.76579589151578, "decode_tok_s_mean_ci95_low": 111.27910969777885, "decode_tok_s_min": 116.95408668502758, "decode_tok_s_p25": 116.95408668502758, "decode_tok_s_p50": 124.46274256962307, "decode_tok_s_p75": 148.65052912929124, "decode_tok_s_p90": 148.65052912929124, "decode_tok_s_p95": 148.65052912929124, "decode_tok_s_p99": 148.65052912929124, "decode_tok_s_samples": 3, "decode_tok_s_stddev": 16.563480891567117, "e2e_max_ms": 61380.194649998884, "e2e_mean_ci95_high_ms": 61094.13163211415, "e2e_mean_ci95_low_ms": 33048.28842254904, "e2e_mean_ms": 47071.210027331595, "e2e_min_ms": 39866.167151998525, "e2e_p25_ms": 39866.167151998525, "e2e_p50_ms": 39967.26827999737, "e2e_p75_ms": 61380.194649998884, "e2e_p90_ms": 61380.194649998884, "e2e_p95_ms": 61380.194649998884, "e2e_p99_ms": 61380.194649998884, "e2e_samples": 3, "e2e_stddev_ms": 12392.047290836063, "effective_prefill_tok_s_max": 5464.71539626543, "effective_prefill_tok_s_mean": 4822.332184640648, "effective_prefill_tok_s_mean_ci95_high": 6075.218420526699, "effective_prefill_tok_s_mean_ci95_low": 3569.4459487545964, "effective_prefill_tok_s_min": 3543.882042901845, "effective_prefill_tok_s_p25": 3543.882042901845, "effective_prefill_tok_s_p50": 5458.399114754668, "effective_prefill_tok_s_p75": 5464.71539626543, "effective_prefill_tok_s_p90": 5464.71539626543, "effective_prefill_tok_s_p95": 5464.71539626543, "effective_prefill_tok_s_p99": 5464.71539626543, "effective_prefill_tok_s_samples": 3, "effective_prefill_tok_s_stddev": 1107.1748044175338, "generation_max_ms": 401.8671029989491, "generation_mean_ci95_high_ms": 413.25219392402005, "generation_mean_ci95_low_ms": 320.06175673989753, "generation_mean_ms": 366.6569753319588, "generation_min_ms": 321.3813160000427, "generation_p25_ms": 321.3813160000427, "generation_p50_ms": 376.7225069968845, "generation_p75_ms": 401.8671029989491, "generation_p90_ms": 401.8671029989491, "generation_p95_ms": 401.8671029989491, "generation_p99_ms": 401.8671029989491, "generation_samples": 3, "generation_stddev_ms": 41.17616632205516, "mean_inter_token_latency_ms_max": 8.550363893594662, "mean_inter_token_latency_ms_mean": 7.770694806180032, "mean_inter_token_latency_ms_mean_ci95_high": 8.834164648115756, "mean_inter_token_latency_ms_mean_ci95_low": 6.707224964244308, "mean_inter_token_latency_ms_min": 6.727187624944365, "mean_inter_token_latency_ms_p25": 6.727187624944365, "mean_inter_token_latency_ms_p50": 8.034532900001068, "mean_inter_token_latency_ms_p75": 8.550363893594662, "mean_inter_token_latency_ms_p90": 8.550363893594662, "mean_inter_token_latency_ms_p95": 8.550363893594662, "mean_inter_token_latency_ms_p99": 8.550363893594662, "mean_inter_token_latency_ms_samples": 3, "mean_inter_token_latency_ms_stddev": 0.9397876523213865, "mean_time_per_output_token_max_ms": 8.550363893594662, "mean_time_per_output_token_mean_ci95_high_ms": 8.834164648115756, "mean_time_per_output_token_mean_ci95_low_ms": 6.707224964244308, "mean_time_per_output_token_mean_ms": 7.770694806180032, "mean_time_per_output_token_min_ms": 6.727187624944365, "mean_time_per_output_token_p25_ms": 6.727187624944365, "mean_time_per_output_token_p50_ms": 8.034532900001068, "mean_time_per_output_token_p75_ms": 8.550363893594662, "mean_time_per_output_token_p90_ms": 8.550363893594662, "mean_time_per_output_token_p95_ms": 8.550363893594662, "mean_time_per_output_token_p99_ms": 8.550363893594662, "mean_time_per_output_token_samples": 3, "mean_time_per_output_token_stddev_ms": 0.9397876523213865, "output_token_sources": { "usage": 3 }, "output_tokens": 146, "prefix_cache_hit_avg": null, "prompt_token_samples": 3, "prompt_tokens": 648302, "prompt_tokens_max": 216101.0, "prompt_tokens_mean": 216100.66666666666, "prompt_tokens_mean_ci95_high": 216101.31999999998, "prompt_tokens_mean_ci95_low": 216100.01333333334, "prompt_tokens_min": 216100.0, "prompt_tokens_p25": 216100.0, "prompt_tokens_p50": 216101.0, "prompt_tokens_p75": 216101.0, "prompt_tokens_p90": 216101.0, "prompt_tokens_p95": 216101.0, "prompt_tokens_p99": 216101.0, "prompt_tokens_samples": 3, "prompt_tokens_stddev": 0.5773502691896257, "reasoning_chunks": 0, "request_canary_passed": 0, "request_canary_samples": 0, "throughput_performance_eligible": true, "throughput_tok_s": 1.0338363102010584, "time_to_first_output_max_ms": 60978.327546999935, "time_to_first_output_mean_ci95_high_ms": 60692.87601818372, "time_to_first_output_mean_ci95_low_ms": 32716.230085815543, "time_to_first_output_mean_ms": 46704.55305199963, "time_to_first_output_min_ms": 39544.78583599848, "time_to_first_output_p25_ms": 39544.78583599848, "time_to_first_output_p50_ms": 39590.545773000485, "time_to_first_output_p75_ms": 60978.327546999935, "time_to_first_output_p90_ms": 60978.327546999935, "time_to_first_output_p95_ms": 60978.327546999935, "time_to_first_output_p99_ms": 60978.327546999935, "time_to_first_output_samples": 3, "time_to_first_output_stddev_ms": 12361.472494955826, "tpot_max_ms": 8.550363893594662, "tpot_mean_ci95_high_ms": 8.834164648115756, "tpot_mean_ci95_low_ms": 6.707224964244308, "tpot_mean_ms": 7.770694806180032, "tpot_min_ms": 6.727187624944365, "tpot_p25_ms": 6.727187624944365, "tpot_p50_ms": 8.034532900001068, "tpot_p75_ms": 8.550363893594662, "tpot_p90_ms": 8.550363893594662, "tpot_p95_ms": 8.550363893594662, "tpot_p99_ms": 8.550363893594662, "tpot_samples": 3, "tpot_stddev_ms": 0.9397876523213865, "ttft_max_ms": 60978.32839699913, "ttft_mean_ci95_high_ms": 60692.87686322718, "ttft_mean_ci95_low_ms": 32716.230427437116, "ttft_mean_ms": 46704.55364533215, "ttft_min_ms": 39544.78621599992, "ttft_p25_ms": 39544.78621599992, "ttft_p50_ms": 39590.5463229974, "ttft_p75_ms": 60978.32839699913, "ttft_p90_ms": 60978.32839699913, "ttft_p95_ms": 60978.32839699913, "ttft_p99_ms": 60978.32839699913, "ttft_samples": 3, "ttft_stddev_ms": 12361.472717392635 }, "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": 23, "controlled_output": null, "decode_tok_s": 116.95408668502758, "e2e_ms": 61380.194649998884, "effective_prefill_tok_s": 3543.882042901845, "generation_ms": 401.8671029989491, "mean_inter_token_latency_ms": 8.550363893594662, "mean_time_per_output_token_ms": 8.550363893594662, "output_token_source": "usage", "output_tokens": 48, "planned_context_tokens": 262144, "prompt_tokens": 216100, "reasoning_chunks": 0, "request_canary": null, "request_index": 0, "time_to_first_output_ms": 60978.327546999935, "tpot_ms": 8.550363893594662, "ttft_ms": 60978.32839699913, "visible_content_truncated": false, "visible_generation_ms": 401.86625299975276 }, { "content_chunks": 23, "controlled_output": null, "decode_tok_s": 148.65052912929124, "e2e_ms": 39967.26827999737, "effective_prefill_tok_s": 5458.399114754668, "generation_ms": 376.7225069968845, "mean_inter_token_latency_ms": 6.727187624944365, "mean_time_per_output_token_ms": 6.727187624944365, "output_token_source": "usage", "output_tokens": 57, "planned_context_tokens": 262144, "prompt_tokens": 216101, "reasoning_chunks": 0, "request_canary": null, "request_index": 1, "time_to_first_output_ms": 39590.545773000485, "tpot_ms": 6.727187624944365, "ttft_ms": 39590.5463229974, "visible_content_truncated": false, "visible_generation_ms": 376.72195699997246 }, { "content_chunks": 19, "controlled_output": null, "decode_tok_s": 124.46274256962307, "e2e_ms": 39866.167151998525, "effective_prefill_tok_s": 5464.71539626543, "generation_ms": 321.3813160000427, "mean_inter_token_latency_ms": 8.034532900001068, "mean_time_per_output_token_ms": 8.034532900001068, "output_token_source": "usage", "output_tokens": 41, "planned_context_tokens": 262144, "prompt_tokens": 216101, "reasoning_chunks": 0, "request_canary": null, "request_index": 2, "time_to_first_output_ms": 39544.78583599848, "tpot_ms": 8.034532900001068, "ttft_ms": 39544.78621599992, "visible_content_truncated": false, "visible_generation_ms": 321.3809359986044 } ], "requests": 3, "response_words": 0, "responses_received": 3, "run_id": "benchmark-20260908T051814Z", "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": "2026-09-08T05:18:14Z", "started_at_unix_ns": 1788844694137699568, "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": 141221.5827200016 }