{ "base_url": "http://127.0.0.1:8001/v1", "clock_domain_id": null, "completed": 3, "concurrency": 1, "configuration_fingerprint": null, "context_distribution": { "4096": 3 }, "context_seed": 0, "context_tokens": 4096, "controlled_output_policy": "observe", "engine": "sglang-glm53-sm120-v0.1.1-rc14", "failed": 0, "failures": [], "finished_at": "2026-09-08T05:17:24Z", "finished_at_unix_ns": 1788844644413036465, "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": 155.0, "completion_tokens_mean": 90.33333333333333, "completion_tokens_mean_ci95_high": 156.24233052727578, "completion_tokens_mean_ci95_low": 24.42433613939089, "completion_tokens_min": 42.0, "completion_tokens_p25": 42.0, "completion_tokens_p50": 74.0, "completion_tokens_p75": 155.0, "completion_tokens_p90": 155.0, "completion_tokens_p95": 155.0, "completion_tokens_p99": 155.0, "completion_tokens_samples": 3, "completion_tokens_stddev": 58.2437407223586, "content_chunks": 105, "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": 155.25176016580093, "decode_tok_s_mean": 150.80335256788774, "decode_tok_s_mean_ci95_high": 155.1911149226435, "decode_tok_s_mean_ci95_low": 146.415590213132, "decode_tok_s_min": 148.1392952650211, "decode_tok_s_p25": 148.1392952650211, "decode_tok_s_p50": 149.01900227284116, "decode_tok_s_p75": 155.25176016580093, "decode_tok_s_p90": 155.25176016580093, "decode_tok_s_p95": 155.25176016580093, "decode_tok_s_p99": 155.25176016580093, "decode_tok_s_samples": 3, "decode_tok_s_stddev": 3.87746292345663, "e2e_max_ms": 1523.209302999021, "e2e_mean_ci95_high_ms": 1519.9965574431399, "e2e_mean_ci95_low_ms": 171.89961255866717, "e2e_mean_ms": 845.9480850009035, "e2e_min_ms": 403.42199800215894, "e2e_p25_ms": 403.42199800215894, "e2e_p50_ms": 611.2129540015303, "e2e_p75_ms": 1523.209302999021, "e2e_p90_ms": 1523.209302999021, "e2e_p95_ms": 1523.209302999021, "e2e_p99_ms": 1523.209302999021, "e2e_samples": 3, "e2e_stddev_ms": 595.6562250174202, "effective_prefill_tok_s_max": 25068.91905192015, "effective_prefill_tok_s_mean": 17933.466061820578, "effective_prefill_tok_s_mean_ci95_high": 30080.534211166756, "effective_prefill_tok_s_mean_ci95_low": 5786.397912474398, "effective_prefill_tok_s_min": 5588.473622771074, "effective_prefill_tok_s_p25": 5588.473622771074, "effective_prefill_tok_s_p50": 23143.00551077051, "effective_prefill_tok_s_p75": 25068.91905192015, "effective_prefill_tok_s_p90": 25068.91905192015, "effective_prefill_tok_s_p95": 25068.91905192015, "effective_prefill_tok_s_p99": 25068.91905192015, "effective_prefill_tok_s_samples": 3, "effective_prefill_tok_s_stddev": 10734.356733504712, "generation_max_ms": 991.9372240001394, "generation_mean_ci95_high_ms": 1002.4818097493642, "generation_mean_ci95_low_ms": 170.75110225412521, "generation_mean_ms": 586.6164560017447, "generation_min_ms": 275.1326970028458, "generation_p25_ms": 275.1326970028458, "generation_p50_ms": 492.77944700224907, "generation_p75_ms": 991.9372240001394, "generation_p90_ms": 991.9372240001394, "generation_p95_ms": 991.9372240001394, "generation_p99_ms": 991.9372240001394, "generation_samples": 3, "generation_stddev_ms": 367.4999601012659, "mean_inter_token_latency_ms_max": 6.750403383592453, "mean_inter_token_latency_ms_mean": 6.634035924741142, "mean_inter_token_latency_ms_mean_ci95_high": 6.82440330173235, "mean_inter_token_latency_ms_mean_ci95_low": 6.443668547749934, "mean_inter_token_latency_ms_min": 6.4411508051957105, "mean_inter_token_latency_ms_p25": 6.4411508051957105, "mean_inter_token_latency_ms_p50": 6.710553585435264, "mean_inter_token_latency_ms_p75": 6.750403383592453, "mean_inter_token_latency_ms_p90": 6.750403383592453, "mean_inter_token_latency_ms_p95": 6.750403383592453, "mean_inter_token_latency_ms_p99": 6.750403383592453, "mean_inter_token_latency_ms_samples": 3, "mean_inter_token_latency_ms_stddev": 0.1682275352308114, "mean_time_per_output_token_max_ms": 6.750403383592453, "mean_time_per_output_token_mean_ci95_high_ms": 6.82440330173235, "mean_time_per_output_token_mean_ci95_low_ms": 6.443668547749934, "mean_time_per_output_token_mean_ms": 6.634035924741142, "mean_time_per_output_token_min_ms": 6.4411508051957105, "mean_time_per_output_token_p25_ms": 6.4411508051957105, "mean_time_per_output_token_p50_ms": 6.710553585435264, "mean_time_per_output_token_p75_ms": 6.750403383592453, "mean_time_per_output_token_p90_ms": 6.750403383592453, "mean_time_per_output_token_p95_ms": 6.750403383592453, "mean_time_per_output_token_p99_ms": 6.750403383592453, "mean_time_per_output_token_samples": 3, "mean_time_per_output_token_stddev_ms": 0.1682275352308114, "output_token_sources": { "usage": 3 }, "output_tokens": 271, "prefix_cache_hit_avg": 0.9915796564499831, "prompt_token_samples": 3, "prompt_tokens": 8907, "prompt_tokens_max": 2969.0, "prompt_tokens_mean": 2969.0, "prompt_tokens_mean_ci95_high": 2969.0, "prompt_tokens_mean_ci95_low": 2969.0, "prompt_tokens_min": 2969.0, "prompt_tokens_p25": 2969.0, "prompt_tokens_p50": 2969.0, "prompt_tokens_p75": 2969.0, "prompt_tokens_p90": 2969.0, "prompt_tokens_p95": 2969.0, "prompt_tokens_p99": 2969.0, "prompt_tokens_samples": 3, "prompt_tokens_stddev": 0.0, "reasoning_chunks": 0, "request_canary_passed": 0, "request_canary_samples": 0, "throughput_performance_eligible": true, "throughput_tok_s": 106.75009323273754, "time_to_first_output_max_ms": 531.2720789988816, "time_to_first_output_mean_ci95_high_ms": 525.8916060085057, "time_to_first_output_mean_ci95_low_ms": -7.22834801018837, "time_to_first_output_mean_ms": 259.3316289991587, "time_to_first_output_min_ms": 118.43350699928124, "time_to_first_output_p25_ms": 118.43350699928124, "time_to_first_output_p50_ms": 128.28930099931313, "time_to_first_output_p75_ms": 531.2720789988816, "time_to_first_output_p90_ms": 531.2720789988816, "time_to_first_output_p95_ms": 531.2720789988816, "time_to_first_output_p99_ms": 531.2720789988816, "time_to_first_output_samples": 3, "time_to_first_output_stddev_ms": 235.5588895125413, "tpot_max_ms": 6.750403383592453, "tpot_mean_ci95_high_ms": 6.82440330173235, "tpot_mean_ci95_low_ms": 6.443668547749934, "tpot_mean_ms": 6.634035924741142, "tpot_min_ms": 6.4411508051957105, "tpot_p25_ms": 6.4411508051957105, "tpot_p50_ms": 6.710553585435264, "tpot_p75_ms": 6.750403383592453, "tpot_p90_ms": 6.750403383592453, "tpot_p95_ms": 6.750403383592453, "tpot_p99_ms": 6.750403383592453, "tpot_samples": 3, "tpot_stddev_ms": 0.1682275352308114, "ttft_max_ms": 531.2728190001508, "ttft_mean_ci95_high_ms": 525.8923436786224, "ttft_mean_ci95_low_ms": -7.228105676723374, "ttft_mean_ms": 259.3321190009495, "ttft_min_ms": 118.43375700118486, "ttft_p25_ms": 118.43375700118486, "ttft_p50_ms": 128.2897810015129, "ttft_p75_ms": 531.2728190001508, "ttft_p90_ms": 531.2728190001508, "ttft_p95_ms": 531.2728190001508, "ttft_p99_ms": 531.2728190001508, "ttft_samples": 3, "ttft_stddev_ms": 235.5591083768901 }, "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": 58, "controlled_output": null, "decode_tok_s": 155.25176016580093, "e2e_ms": 1523.209302999021, "effective_prefill_tok_s": 5588.473622771074, "generation_ms": 991.9372240001394, "mean_inter_token_latency_ms": 6.4411508051957105, "mean_time_per_output_token_ms": 6.4411508051957105, "output_token_source": "usage", "output_tokens": 155, "planned_context_tokens": 4096, "prompt_tokens": 2969, "reasoning_chunks": 0, "request_canary": null, "request_index": 0, "time_to_first_output_ms": 531.2720789988816, "tpot_ms": 6.4411508051957105, "ttft_ms": 531.2728190001508, "visible_content_truncated": false, "visible_generation_ms": 991.9364839988702 }, { "content_chunks": 17, "controlled_output": null, "decode_tok_s": 149.01900227284116, "e2e_ms": 403.42199800215894, "effective_prefill_tok_s": 23143.00551077051, "generation_ms": 275.1326970028458, "mean_inter_token_latency_ms": 6.710553585435264, "mean_time_per_output_token_ms": 6.710553585435264, "output_token_source": "usage", "output_tokens": 42, "planned_context_tokens": 4096, "prompt_tokens": 2969, "reasoning_chunks": 0, "request_canary": null, "request_index": 1, "time_to_first_output_ms": 128.28930099931313, "tpot_ms": 6.710553585435264, "ttft_ms": 128.2897810015129, "visible_content_truncated": false, "visible_generation_ms": 275.13221700064605 }, { "content_chunks": 30, "controlled_output": null, "decode_tok_s": 148.1392952650211, "e2e_ms": 611.2129540015303, "effective_prefill_tok_s": 25068.91905192015, "generation_ms": 492.77944700224907, "mean_inter_token_latency_ms": 6.750403383592453, "mean_time_per_output_token_ms": 6.750403383592453, "output_token_source": "usage", "output_tokens": 74, "planned_context_tokens": 4096, "prompt_tokens": 2969, "reasoning_chunks": 0, "request_canary": null, "request_index": 2, "time_to_first_output_ms": 118.43350699928124, "tpot_ms": 6.750403383592453, "ttft_ms": 118.43375700118486, "visible_content_truncated": false, "visible_generation_ms": 492.77919700034545 } ], "requests": 3, "response_words": 0, "responses_received": 3, "run_id": "benchmark-20260908T051721Z", "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": "2026-09-08T05:17:21Z", "started_at_unix_ns": 1788844641874394186, "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": 2538.639468999463 }