{ "base_url": "http://127.0.0.1:30000/v1", "clock_domain_id": null, "completed": 6, "concurrency": 8, "configuration_fingerprint": null, "context_distribution": { "32768": 8 }, "context_seed": 8801, "context_tokens": 32768, "controlled_output_policy": "strict", "engine": "vllm-exl3-v84", "failed": 2, "failures": [ { "elapsed_ms": 30027.30214100302, "error_message": "HTTP Error 503: Service Unavailable", "error_message_truncated": false, "error_type": "HTTPError", "failure_class": "request_error", "performance_eligible": false, "planned_context_tokens": 32768, "request_index": 3, "response_received": false }, { "elapsed_ms": 30024.66827900207, "error_message": "HTTP Error 503: Service Unavailable", "error_message_truncated": false, "error_type": "HTTPError", "failure_class": "request_error", "performance_eligible": false, "planned_context_tokens": 32768, "request_index": 6, "response_received": false } ], "finished_at": "2026-09-24T02:06:34Z", "finished_at_unix_ns": 1790215594707462271, "gpu": "dual-pro6000-maxq", "manifest": null, "max_context_tokens": 327680, "max_tokens": 8192, "measurement_protocol": "capacity-v4-reasoning", "metric_population": { "aggregate_throughput_eligible": false, "attempted_requests": 8, "excluded_requests": 2, "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": 6, "e2e": 6, "effective_prefill_tok_s": 6, "tpot": 6, "ttft": 6 }, "performance_eligible_requests": 6, "successful_requests": 6 }, "metrics": { "completion_tokens_max": 390.0, "completion_tokens_mean": 238.66666666666666, "completion_tokens_mean_ci95_high": 333.40135171449435, "completion_tokens_mean_ci95_low": 143.93198161883896, "completion_tokens_min": 132.0, "completion_tokens_p25": 134.0, "completion_tokens_p50": 134.0, "completion_tokens_p75": 334.0, "completion_tokens_p90": 390.0, "completion_tokens_p95": 390.0, "completion_tokens_p99": 390.0, "completion_tokens_samples": 6, "completion_tokens_stddev": 118.39369352573922, "content_chunks": 162, "content_chunks_s": null, "controlled_output_exact_adherence": 6, "controlled_output_exact_adherence_rate": 1.0, "controlled_output_samples": 6, "controlled_output_unobservable": 0, "decode_tok_s_max": 63.902855316828585, "decode_tok_s_mean": 24.210259423576687, "decode_tok_s_mean_ci95_high": 39.97820489226637, "decode_tok_s_mean_ci95_low": 8.442313954887007, "decode_tok_s_min": 12.573786422333889, "decode_tok_s_p25": 13.279193076834765, "decode_tok_s_p50": 15.55233071263643, "decode_tok_s_p75": 20.42072540515593, "decode_tok_s_p90": 63.902855316828585, "decode_tok_s_p95": 63.902855316828585, "decode_tok_s_p99": 63.902855316828585, "decode_tok_s_samples": 6, "decode_tok_s_stddev": 19.705826882816257, "e2e_max_ms": 36488.99111000355, "e2e_mean_ci95_high_ms": 35865.06520565592, "e2e_mean_ci95_low_ms": 25444.95063201086, "e2e_mean_ms": 30655.007918833388, "e2e_min_ms": 21194.342918002803, "e2e_p25_ms": 23622.230105997005, "e2e_p50_ms": 33436.57934899966, "e2e_p75_ms": 34919.84558299737, "e2e_p90_ms": 36488.99111000355, "e2e_p95_ms": 36488.99111000355, "e2e_p99_ms": 36488.99111000355, "e2e_samples": 6, "e2e_stddev_ms": 6511.2152466247635, "effective_prefill_tok_s_max": 4494.216702717941, "effective_prefill_tok_s_mean": 1935.2871690141565, "effective_prefill_tok_s_mean_ci95_high": 3035.2377359136244, "effective_prefill_tok_s_mean_ci95_low": 835.3366021146887, "effective_prefill_tok_s_min": 808.3648341033226, "effective_prefill_tok_s_p25": 949.4705008726669, "effective_prefill_tok_s_p50": 1314.4738155815646, "effective_prefill_tok_s_p75": 2381.606926023346, "effective_prefill_tok_s_p90": 4494.216702717941, "effective_prefill_tok_s_p95": 4494.216702717941, "effective_prefill_tok_s_p99": 4494.216702717941, "effective_prefill_tok_s_samples": 6, "effective_prefill_tok_s_stddev": 1374.6518526473408, "generation_max_ms": 29293.948642000032, "generation_mean_ci95_high_ms": 19663.538159651398, "generation_mean_ci95_low_ms": 5452.652604019203, "generation_mean_ms": 12558.095381835301, "generation_min_ms": 5211.034755004221, "generation_p25_ms": 6809.106482003699, "generation_p50_ms": 8423.174791001657, "generation_p75_ms": 15033.746054999938, "generation_p90_ms": 29293.948642000032, "generation_p95_ms": 29293.948642000032, "generation_p99_ms": 29293.948642000032, "generation_samples": 6, "generation_stddev_ms": 8879.95367458844, "mean_inter_token_latency_ms_max": 79.53053809024256, "mean_inter_token_latency_ms_mean": 55.825043680022986, "mean_inter_token_latency_ms_mean_ci95_high": 74.40480832613858, "mean_inter_token_latency_ms_mean_ci95_low": 37.245279033907394, "mean_inter_token_latency_ms_min": 15.648753018030693, "mean_inter_token_latency_ms_p25": 48.969856856677325, "mean_inter_token_latency_ms_p50": 51.196289338373674, "mean_inter_token_latency_ms_p75": 75.30578057069417, "mean_inter_token_latency_ms_p90": 79.53053809024256, "mean_inter_token_latency_ms_p95": 79.53053809024256, "mean_inter_token_latency_ms_p99": 79.53053809024256, "mean_inter_token_latency_ms_samples": 6, "mean_inter_token_latency_ms_stddev": 23.219868838768196, "mean_time_per_output_token_max_ms": 79.53053809024256, "mean_time_per_output_token_mean_ci95_high_ms": 74.40480832613858, "mean_time_per_output_token_mean_ci95_low_ms": 37.245279033907394, "mean_time_per_output_token_mean_ms": 55.825043680022986, "mean_time_per_output_token_min_ms": 15.648753018030693, "mean_time_per_output_token_p25_ms": 48.969856856677325, "mean_time_per_output_token_p50_ms": 51.196289338373674, "mean_time_per_output_token_p75_ms": 75.30578057069417, "mean_time_per_output_token_p90_ms": 79.53053809024256, "mean_time_per_output_token_p95_ms": 79.53053809024256, "mean_time_per_output_token_p99_ms": 79.53053809024256, "mean_time_per_output_token_samples": 6, "mean_time_per_output_token_stddev_ms": 23.219868838768196, "output_token_sources": { "usage": 6 }, "output_tokens": 1432, "prefix_cache_hit_avg": null, "prompt_token_samples": 6, "prompt_tokens": 151703, "prompt_tokens_max": 25285.0, "prompt_tokens_mean": 25283.833333333332, "prompt_tokens_mean_ci95_high": 25284.768764309465, "prompt_tokens_mean_ci95_low": 25282.8979023572, "prompt_tokens_min": 25282.0, "prompt_tokens_p25": 25283.0, "prompt_tokens_p50": 25284.0, "prompt_tokens_p75": 25285.0, "prompt_tokens_p90": 25285.0, "prompt_tokens_p95": 25285.0, "prompt_tokens_p99": 25285.0, "prompt_tokens_samples": 6, "prompt_tokens_stddev": 1.1690451944500122, "reasoning_chunks": 1154, "request_canary_passed": 6, "request_canary_samples": 6, "throughput_performance_eligible": false, "throughput_tok_s": null, "time_to_first_output_max_ms": 31277.956354999333, "time_to_first_output_mean_ci95_high_ms": 25835.647026906794, "time_to_first_output_mean_ci95_low_ms": 10358.17804708938, "time_to_first_output_mean_ms": 18096.912536998087, "time_to_first_output_min_ms": 5625.896940997336, "time_to_first_output_p25_ms": 10616.781352000544, "time_to_first_output_p50_ms": 15199.055314995348, "time_to_first_output_p75_ms": 26627.47286699596, "time_to_first_output_p90_ms": 31277.956354999333, "time_to_first_output_p95_ms": 31277.956354999333, "time_to_first_output_p99_ms": 31277.956354999333, "time_to_first_output_samples": 6, "time_to_first_output_stddev_ms": 9671.403446507034, "tpot_max_ms": 79.53053809024256, "tpot_mean_ci95_high_ms": 74.40480832613858, "tpot_mean_ci95_low_ms": 37.245279033907394, "tpot_mean_ms": 55.825043680022986, "tpot_min_ms": 15.648753018030693, "tpot_p25_ms": 48.969856856677325, "tpot_p50_ms": 51.196289338373674, "tpot_p75_ms": 75.30578057069417, "tpot_p90_ms": 79.53053809024256, "tpot_p95_ms": 79.53053809024256, "tpot_p99_ms": 79.53053809024256, "tpot_samples": 6, "tpot_stddev_ms": 23.219868838768196, "ttft_max_ms": 36140.67260199954, "ttft_mean_ci95_high_ms": 35498.45600829646, "ttft_mean_ci95_low_ms": 24174.380874032588, "ttft_mean_ms": 29836.418441164522, "ttft_min_ms": 20711.10806700017, "ttft_p25_ms": 20894.526296993718, "ttft_p50_ms": 32943.79636499798, "ttft_p75_ms": 34529.202424993855, "ttft_p90_ms": 36140.67260199954, "ttft_p95_ms": 36140.67260199954, "ttft_p99_ms": 36140.67260199954, "ttft_samples": 6, "ttft_stddev_ms": 7076.0729305830055 }, "model": "llm.primary", "performance_eligible": false, "prompt_cache_mode": "unique", "request_canaries": true, "request_timings": [ { "content_chunks": 27, "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": 20.42072540515593, "e2e_ms": 34268.05844699993, "effective_prefill_tok_s": 1314.4738155815646, "generation_ms": 15033.746054999938, "mean_inter_token_latency_ms": 48.969856856677325, "mean_time_per_output_token_ms": 48.969856856677325, "output_token_source": "usage", "output_tokens": 308, "planned_context_tokens": 32768, "prompt_tokens": 25283, "reasoning_chunks": 257, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_8801_00000", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 0, "time_to_first_output_ms": 19234.312391999993, "tpot_ms": 48.969856856677325, "ttft_ms": 33799.204891001864, "visible_content_truncated": false, "visible_generation_ms": 468.85355599806644 }, { "content_chunks": 27, "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": 13.279193076834765, "e2e_ms": 34919.84558299737, "effective_prefill_tok_s": 4494.216702717941, "generation_ms": 29293.948642000032, "mean_inter_token_latency_ms": 75.30578057069417, "mean_time_per_output_token_ms": 75.30578057069417, "output_token_source": "usage", "output_tokens": 390, "planned_context_tokens": 32768, "prompt_tokens": 25284, "reasoning_chunks": 318, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_8801_00001", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 1, "time_to_first_output_ms": 5625.896940997336, "tpot_ms": 75.30578057069417, "ttft_ms": 34529.202424993855, "visible_content_truncated": false, "visible_generation_ms": 390.64315800351324 }, { "content_chunks": 27, "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": 15.55233071263643, "e2e_ms": 23622.230105997005, "effective_prefill_tok_s": 1663.5902347860979, "generation_ms": 8423.174791001657, "mean_inter_token_latency_ms": 64.29904420611952, "mean_time_per_output_token_ms": 64.29904420611952, "output_token_source": "usage", "output_tokens": 132, "planned_context_tokens": 32768, "prompt_tokens": 25285, "reasoning_chunks": 97, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_8801_00002", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 2, "time_to_first_output_ms": 15199.055314995348, "tpot_ms": 64.29904420611952, "ttft_ms": 20894.526296993718, "visible_content_truncated": false, "visible_generation_ms": 2727.7038090032875 }, { "content_chunks": 27, "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": 12.573786422333889, "e2e_ms": 21194.342918002803, "effective_prefill_tok_s": 2381.606926023346, "generation_ms": 10577.56156600226, "mean_inter_token_latency_ms": 79.53053809024256, "mean_time_per_output_token_ms": 79.53053809024256, "output_token_source": "usage", "output_tokens": 134, "planned_context_tokens": 32768, "prompt_tokens": 25285, "reasoning_chunks": 99, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_8801_00004", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 4, "time_to_first_output_ms": 10616.781352000544, "tpot_ms": 79.53053809024256, "ttft_ms": 20711.10806700017, "visible_content_truncated": false, "visible_generation_ms": 483.23485100263497 }, { "content_chunks": 27, "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": 63.902855316828585, "e2e_ms": 36488.99111000355, "effective_prefill_tok_s": 808.3648341033226, "generation_ms": 5211.034755004221, "mean_inter_token_latency_ms": 15.648753018030693, "mean_time_per_output_token_ms": 15.648753018030693, "output_token_source": "usage", "output_tokens": 334, "planned_context_tokens": 32768, "prompt_tokens": 25284, "reasoning_chunks": 284, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_8801_00005", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 5, "time_to_first_output_ms": 31277.956354999333, "tpot_ms": 15.648753018030693, "ttft_ms": 36140.67260199954, "visible_content_truncated": false, "visible_generation_ms": 348.318508004013 }, { "content_chunks": 27, "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": 19.532665607670513, "e2e_ms": 33436.57934899966, "effective_prefill_tok_s": 949.4705008726669, "generation_ms": 6809.106482003699, "mean_inter_token_latency_ms": 51.196289338373674, "mean_time_per_output_token_ms": 51.196289338373674, "output_token_source": "usage", "output_tokens": 134, "planned_context_tokens": 32768, "prompt_tokens": 25282, "reasoning_chunks": 99, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_8801_00007", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 7, "time_to_first_output_ms": 26627.47286699596, "tpot_ms": 51.196289338373674, "ttft_ms": 32943.79636499798, "visible_content_truncated": false, "visible_generation_ms": 492.7829840016784 } ], "requests": 8, "response_words": 16, "responses_received": 6, "run_id": "benchmark-20260924T020558Z", "sampling": { "temperature": { "effective_request": 0.6, "requested": 0.6, "sent": true }, "top_p": { "effective_request": 0.95, "requested": 0.95, "sent": true } }, "schema": "anvil-serving.benchmark/v1", "serve_flags": { "no_thinking": false, "reasoning_effort": "max", "shared_prefix_burst": false, "thinking_mode": "default" }, "shared_prefix_tokens": 0, "source_recipe": null, "started_at": "2026-09-24T02:05:58Z", "started_at_unix_ns": 1790215558212755525, "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": 36494.70331599878 }