{ "base_url": "http://127.0.0.1:/v1", "clock_domain_id": null, "completed": 8, "concurrency": 4, "configuration_fingerprint": null, "context_distribution": { "32768": 8 }, "context_seed": 0, "context_tokens": 32768, "controlled_output_policy": "strict", "engine": "vllm-glm53-exl3-v84", "failed": 0, "failures": [], "finished_at": "2026-09-22T02:19:22Z", "finished_at_unix_ns": 1790043562853709461, "gpu": "rtx-pro6000-maxq-tp2", "manifest": null, "max_context_tokens": 327680, "max_tokens": 2048, "measurement_protocol": "capacity-v4-reasoning", "metric_population": { "aggregate_throughput_eligible": true, "attempted_requests": 8, "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": 8, "e2e": 8, "effective_prefill_tok_s": 8, "tpot": 8, "ttft": 8 }, "performance_eligible_requests": 8, "successful_requests": 8 }, "metrics": { "completion_tokens_max": 179.0, "completion_tokens_mean": 147.625, "completion_tokens_mean_ci95_high": 169.22930107640605, "completion_tokens_mean_ci95_low": 126.02069892359394, "completion_tokens_min": 116.0, "completion_tokens_p25": 118.0, "completion_tokens_p50": 120.0, "completion_tokens_p75": 176.0, "completion_tokens_p90": 179.0, "completion_tokens_p95": 179.0, "completion_tokens_p99": 179.0, "completion_tokens_samples": 8, "completion_tokens_stddev": 31.176628150862353, "content_chunks": 200, "content_chunks_s": null, "controlled_output_exact_adherence": 8, "controlled_output_exact_adherence_rate": 1.0, "controlled_output_samples": 8, "controlled_output_unobservable": 0, "decode_tok_s_max": 53.92697399958375, "decode_tok_s_mean": 15.75173716494867, "decode_tok_s_mean_ci95_high": 26.794264373053046, "decode_tok_s_mean_ci95_low": 4.709209956844294, "decode_tok_s_min": 7.18657894159897, "decode_tok_s_p25": 7.58177970549957, "decode_tok_s_p50": 8.901533807301083, "decode_tok_s_p75": 11.736464933732776, "decode_tok_s_p90": 53.92697399958375, "decode_tok_s_p95": 53.92697399958375, "decode_tok_s_p99": 53.92697399958375, "decode_tok_s_samples": 8, "decode_tok_s_stddev": 15.935195653648082, "e2e_max_ms": 39395.753278862685, "e2e_mean_ci95_high_ms": 30548.613902726873, "e2e_mean_ci95_low_ms": 17068.72414713892, "e2e_mean_ms": 23808.669024932897, "e2e_min_ms": 8321.754673030227, "e2e_p25_ms": 15148.439358919859, "e2e_p50_ms": 21992.66061000526, "e2e_p75_ms": 26858.828680822626, "e2e_p90_ms": 39395.753278862685, "e2e_p95_ms": 39395.753278862685, "e2e_p99_ms": 39395.753278862685, "e2e_samples": 8, "e2e_stddev_ms": 9726.246383493175, "effective_prefill_tok_s_max": 4730.791606332443, "effective_prefill_tok_s_mean": 3326.630559967605, "effective_prefill_tok_s_mean_ci95_high": 4299.294931210413, "effective_prefill_tok_s_mean_ci95_low": 2353.9661887247976, "effective_prefill_tok_s_min": 1237.6867521606557, "effective_prefill_tok_s_p25": 1546.4651545329775, "effective_prefill_tok_s_p50": 4109.455008734245, "effective_prefill_tok_s_p75": 4219.937541823812, "effective_prefill_tok_s_p90": 4730.791606332443, "effective_prefill_tok_s_p95": 4730.791606332443, "effective_prefill_tok_s_p99": 4730.791606332443, "effective_prefill_tok_s_samples": 8, "effective_prefill_tok_s_stddev": 1403.6276994374266, "generation_max_ms": 23047.502608038485, "generation_mean_ci95_high_ms": 18626.466435568003, "generation_mean_ci95_low_ms": 9658.848798355026, "generation_mean_ms": 14142.657616961515, "generation_min_ms": 2169.600690016523, "generation_p25_ms": 9012.079263105989, "generation_p50_ms": 14910.793069982901, "generation_p75_ms": 16002.05061887391, "generation_p90_ms": 23047.502608038485, "generation_p95_ms": 23047.502608038485, "generation_p99_ms": 23047.502608038485, "generation_samples": 8, "generation_stddev_ms": 6470.472696287125, "mean_inter_token_latency_ms_max": 139.14826625107747, "mean_inter_token_latency_ms_mean": 96.7188550373315, "mean_inter_token_latency_ms_mean_ci95_high": 126.39657850259249, "mean_inter_token_latency_ms_mean_ci95_low": 67.0411315720705, "mean_inter_token_latency_ms_min": 18.543595641166863, "mean_inter_token_latency_ms_p25": 51.497595789177076, "mean_inter_token_latency_ms_p50": 103.42149204034514, "mean_inter_token_latency_ms_p75": 131.70001490307706, "mean_inter_token_latency_ms_p90": 139.14826625107747, "mean_inter_token_latency_ms_p95": 139.14826625107747, "mean_inter_token_latency_ms_p99": 139.14826625107747, "mean_inter_token_latency_ms_samples": 8, "mean_inter_token_latency_ms_stddev": 42.82718267850035, "mean_time_per_output_token_max_ms": 139.14826625107747, "mean_time_per_output_token_mean_ci95_high_ms": 126.39657850259249, "mean_time_per_output_token_mean_ci95_low_ms": 67.0411315720705, "mean_time_per_output_token_mean_ms": 96.7188550373315, "mean_time_per_output_token_min_ms": 18.543595641166863, "mean_time_per_output_token_p25_ms": 51.497595789177076, "mean_time_per_output_token_p50_ms": 103.42149204034514, "mean_time_per_output_token_p75_ms": 131.70001490307706, "mean_time_per_output_token_p90_ms": 139.14826625107747, "mean_time_per_output_token_p95_ms": 139.14826625107747, "mean_time_per_output_token_p99_ms": 139.14826625107747, "mean_time_per_output_token_samples": 8, "mean_time_per_output_token_stddev_ms": 42.82718267850035, "output_token_sources": { "usage": 8 }, "output_tokens": 1181, "prefix_cache_hit_avg": null, "prompt_token_samples": 8, "prompt_tokens": 202253, "prompt_tokens_max": 25283.0, "prompt_tokens_mean": 25281.625, "prompt_tokens_mean_ci95_high": 25282.668563606112, "prompt_tokens_mean_ci95_low": 25280.581436393888, "prompt_tokens_min": 25279.0, "prompt_tokens_p25": 25280.0, "prompt_tokens_p50": 25282.0, "prompt_tokens_p75": 25283.0, "prompt_tokens_p90": 25283.0, "prompt_tokens_p95": 25283.0, "prompt_tokens_p99": 25283.0, "prompt_tokens_samples": 8, "prompt_tokens_stddev": 1.5059406173077154, "reasoning_chunks": 872, "request_canary_passed": 8, "request_canary_samples": 8, "throughput_performance_eligible": true, "throughput_tok_s": 24.663643174204225, "time_to_first_output_max_ms": 20426.00840306841, "time_to_first_output_mean_ci95_high_ms": 13679.950989022787, "time_to_first_output_mean_ci95_low_ms": 5652.071826919977, "time_to_first_output_mean_ms": 9666.011407971382, "time_to_first_output_min_ms": 5344.348706072196, "time_to_first_output_p25_ms": 5767.054029973224, "time_to_first_output_p50_ms": 6136.36009581387, "time_to_first_output_p75_ms": 11163.305383874103, "time_to_first_output_p90_ms": 20426.00840306841, "time_to_first_output_p95_ms": 20426.00840306841, "time_to_first_output_p99_ms": 20426.00840306841, "time_to_first_output_samples": 8, "time_to_first_output_stddev_ms": 5792.416116397018, "tpot_max_ms": 139.14826625107747, "tpot_mean_ci95_high_ms": 126.39657850259249, "tpot_mean_ci95_low_ms": 67.0411315720705, "tpot_mean_ms": 96.7188550373315, "tpot_min_ms": 18.543595641166863, "tpot_p25_ms": 51.497595789177076, "tpot_p50_ms": 103.42149204034514, "tpot_p75_ms": 131.70001490307706, "tpot_p90_ms": 139.14826625107747, "tpot_p95_ms": 139.14826625107747, "tpot_p99_ms": 139.14826625107747, "tpot_samples": 8, "tpot_stddev_ms": 42.82718267850035, "ttft_max_ms": 38919.218197930604, "ttft_mean_ci95_high_ms": 28100.986697658802, "ttft_mean_ci95_low_ms": 15711.936574015861, "ttft_mean_ms": 21906.46163583733, "ttft_min_ms": 7897.960317088291, "ttft_p25_ms": 14784.396077971905, "ttft_p50_ms": 21523.919400991872, "ttft_p75_ms": 24864.050650037825, "ttft_p90_ms": 38919.218197930604, "ttft_p95_ms": 38919.218197930604, "ttft_p99_ms": 38919.218197930604, "ttft_samples": 8, "ttft_stddev_ms": 8939.164647844858 }, "model": "glm53-flash-exl3-4bpw-v84-fp8-tp2-c4-327k-r10-vision8-apc-nospec", "performance_eligible": true, "prompt_cache_mode": "unique", "request_canaries": true, "request_timings": [ { "content_chunks": 25, "controlled_output": { "capture_complete": true, "exact_adherence": true, "observed_code_words": 16, "observed_extra_words": 0, "passed": true, "policy": "strict", "requested_words": 16 }, "decode_tok_s": 7.58177970549957, "e2e_ms": 26858.828680822626, "effective_prefill_tok_s": 2264.4726746001816, "generation_ms": 15695.523296948522, "mean_inter_token_latency_ms": 131.89515375586993, "mean_time_per_output_token_ms": 131.89515375586993, "output_token_source": "usage", "output_tokens": 120, "planned_context_tokens": 32768, "prompt_tokens": 25279, "reasoning_chunks": 90, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_0_00000", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 0, "time_to_first_output_ms": 11163.305383874103, "tpot_ms": 131.89515375586993, "ttft_ms": 21804.419986903667, "visible_content_truncated": false, "visible_generation_ms": 5054.408693918958 }, { "content_chunks": 25, "controlled_output": { "capture_complete": true, "exact_adherence": true, "observed_code_words": 16, "observed_extra_words": 0, "passed": true, "policy": "strict", "requested_words": 16 }, "decode_tok_s": 7.18657894159897, "e2e_ms": 21992.66061000526, "effective_prefill_tok_s": 4219.937541823812, "generation_ms": 16002.05061887391, "mean_inter_token_latency_ms": 139.14826625107747, "mean_time_per_output_token_ms": 139.14826625107747, "output_token_source": "usage", "output_tokens": 116, "planned_context_tokens": 32768, "prompt_tokens": 25280, "reasoning_chunks": 86, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_0_00001", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 1, "time_to_first_output_ms": 5990.60999113135, "tpot_ms": 139.14826625107747, "ttft_ms": 21523.919400991872, "visible_content_truncated": false, "visible_generation_ms": 468.74120901338756 }, { "content_chunks": 25, "controlled_output": { "capture_complete": true, "exact_adherence": true, "observed_code_words": 16, "observed_extra_words": 0, "passed": true, "policy": "strict", "requested_words": 16 }, "decode_tok_s": 9.669170114176035, "e2e_ms": 32733.16595586948, "effective_prefill_tok_s": 1237.6867521606557, "generation_ms": 12307.157552801073, "mean_inter_token_latency_ms": 103.42149204034514, "mean_time_per_output_token_ms": 103.42149204034514, "output_token_source": "usage", "output_tokens": 120, "planned_context_tokens": 32768, "prompt_tokens": 25281, "reasoning_chunks": 90, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_0_00002", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 2, "time_to_first_output_ms": 20426.00840306841, "tpot_ms": 103.42149204034514, "ttft_ms": 25249.584707897156, "visible_content_truncated": false, "visible_generation_ms": 7483.5812479723245 }, { "content_chunks": 25, "controlled_output": { "capture_complete": true, "exact_adherence": true, "observed_code_words": 16, "observed_extra_words": 0, "passed": true, "policy": "strict", "requested_words": 16 }, "decode_tok_s": 7.5930135675074695, "e2e_ms": 39395.753278862685, "effective_prefill_tok_s": 1546.4651545329775, "generation_ms": 23047.502608038485, "mean_inter_token_latency_ms": 131.70001490307706, "mean_time_per_output_token_ms": 131.70001490307706, "output_token_source": "usage", "output_tokens": 176, "planned_context_tokens": 32768, "prompt_tokens": 25282, "reasoning_chunks": 129, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_0_00003", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 3, "time_to_first_output_ms": 16348.2506708242, "tpot_ms": 131.70001490307706, "ttft_ms": 38919.218197930604, "visible_content_truncated": false, "visible_generation_ms": 476.53508093208075 }, { "content_chunks": 25, "controlled_output": { "capture_complete": true, "exact_adherence": true, "observed_code_words": 16, "observed_extra_words": 0, "passed": true, "policy": "strict", "requested_words": 16 }, "decode_tok_s": 8.901533807301083, "e2e_ms": 25340.90254199691, "effective_prefill_tok_s": 4730.791606332443, "generation_ms": 19996.553835924715, "mean_inter_token_latency_ms": 112.34019008946468, "mean_time_per_output_token_ms": 112.34019008946468, "output_token_source": "usage", "output_tokens": 179, "planned_context_tokens": 32768, "prompt_tokens": 25283, "reasoning_chunks": 131, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_0_00004", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 4, "time_to_first_output_ms": 5344.348706072196, "tpot_ms": 112.34019008946468, "ttft_ms": 24864.050650037825, "visible_content_truncated": false, "visible_generation_ms": 476.85189195908606 }, { "content_chunks": 25, "controlled_output": { "capture_complete": true, "exact_adherence": true, "observed_code_words": 16, "observed_extra_words": 0, "passed": true, "policy": "strict", "requested_words": 16 }, "decode_tok_s": 11.736464933732776, "e2e_ms": 20677.847099956125, "effective_prefill_tok_s": 4384.040771700103, "generation_ms": 14910.793069982901, "mean_inter_token_latency_ms": 85.20453182847372, "mean_time_per_output_token_ms": 85.20453182847372, "output_token_source": "usage", "output_tokens": 176, "planned_context_tokens": 32768, "prompt_tokens": 25283, "reasoning_chunks": 129, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_0_00005", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 5, "time_to_first_output_ms": 5767.054029973224, "tpot_ms": 85.20453182847372, "ttft_ms": 20208.14374787733, "visible_content_truncated": false, "visible_generation_ms": 469.70335207879543 }, { "content_chunks": 25, "controlled_output": { "capture_complete": true, "exact_adherence": true, "observed_code_words": 16, "observed_extra_words": 0, "passed": true, "policy": "strict", "requested_words": 16 }, "decode_tok_s": 19.4183822501897, "e2e_ms": 15148.439358919859, "effective_prefill_tok_s": 4120.194969856425, "generation_ms": 9012.079263105989, "mean_inter_token_latency_ms": 51.497595789177076, "mean_time_per_output_token_ms": 51.497595789177076, "output_token_source": "usage", "output_tokens": 176, "planned_context_tokens": 32768, "prompt_tokens": 25283, "reasoning_chunks": 129, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_0_00006", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 6, "time_to_first_output_ms": 6136.36009581387, "tpot_ms": 51.497595789177076, "ttft_ms": 14784.396077971905, "visible_content_truncated": false, "visible_generation_ms": 364.04328094795346 }, { "content_chunks": 25, "controlled_output": { "capture_complete": true, "exact_adherence": true, "observed_code_words": 16, "observed_extra_words": 0, "passed": true, "policy": "strict", "requested_words": 16 }, "decode_tok_s": 53.92697399958375, "e2e_ms": 8321.754673030227, "effective_prefill_tok_s": 4109.455008734245, "generation_ms": 2169.600690016523, "mean_inter_token_latency_ms": 18.543595641166863, "mean_time_per_output_token_ms": 18.543595641166863, "output_token_source": "usage", "output_tokens": 118, "planned_context_tokens": 32768, "prompt_tokens": 25282, "reasoning_chunks": 88, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_0_00007", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 7, "time_to_first_output_ms": 6152.153983013704, "tpot_ms": 18.543595641166863, "ttft_ms": 7897.960317088291, "visible_content_truncated": false, "visible_generation_ms": 423.7943559419364 } ], "requests": 8, "response_words": 16, "responses_received": 8, "run_id": "benchmark-20260922T021834Z", "sampling": { "temperature": { "effective_request": 0.0, "requested": null, "sent": true }, "top_p": { "effective_request": null, "requested": null, "sent": false } }, "schema": "anvil-serving.benchmark/v1", "serve_flags": { "no_thinking": false, "reasoning_effort": null, "shared_prefix_burst": false, "thinking_mode": "enabled" }, "shared_prefix_tokens": 0, "source_recipe": null, "started_at": "2026-09-22T02:18:34Z", "started_at_unix_ns": 1790043514969457888, "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": 47884.24774305895 }