{ "base_url": "[operator identity redacted]", "clock_domain_id": null, "completed": 6, "concurrency": 1, "configuration_fingerprint": null, "context_distribution": { "65536": 6 }, "context_seed": 1717, "context_tokens": 65536, "controlled_output_policy": "observe", "engine": "ninfer", "failed": 0, "failures": [], "finished_at": "2026-10-05T11:44:30Z", "finished_at_unix_ns": 1791200670200584800, "gpu": "rtx5090", "manifest": null, "max_context_tokens": 262144, "max_tokens": 8192, "measurement_protocol": "capacity-v4-reasoning", "metric_population": { "aggregate_throughput_eligible": true, "attempted_requests": 6, "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": 6, "e2e": 6, "effective_prefill_tok_s": 6, "tpot": 6, "ttft": 6 }, "performance_eligible_requests": 6, "successful_requests": 6 }, "metrics": { "completion_tokens_max": 421.0, "completion_tokens_mean": 327.6666666666667, "completion_tokens_mean_ci95_high": 382.55055541777875, "completion_tokens_mean_ci95_low": 272.7827779155546, "completion_tokens_min": 250.0, "completion_tokens_p25": 254.0, "completion_tokens_p50": 317.0, "completion_tokens_p75": 383.0, "completion_tokens_p90": 421.0, "completion_tokens_p95": 421.0, "completion_tokens_p99": 421.0, "completion_tokens_samples": 6, "completion_tokens_stddev": 68.59057272443981, "content_chunks": 55, "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": 273.90026875134754, "decode_tok_s_mean": 234.7498725319921, "decode_tok_s_mean_ci95_high": 252.72124715478915, "decode_tok_s_mean_ci95_low": 216.77849790919507, "decode_tok_s_min": 214.44920928196512, "decode_tok_s_p25": 216.04551844801497, "decode_tok_s_p50": 222.82018666318382, "decode_tok_s_p75": 243.44193243928243, "decode_tok_s_p90": 273.90026875134754, "decode_tok_s_p95": 273.90026875134754, "decode_tok_s_p99": 273.90026875134754, "decode_tok_s_samples": 6, "decode_tok_s_stddev": 22.4595396950282, "e2e_max_ms": 9968.378499997925, "e2e_mean_ci95_high_ms": 7683.649644394931, "e2e_mean_ci95_low_ms": 781.0399889375244, "e2e_mean_ms": 4232.344816666227, "e2e_min_ms": 1211.8689000017184, "e2e_p25_ms": 1260.367300001235, "e2e_p50_ms": 1656.939099997544, "e2e_p75_ms": 9618.114900000364, "e2e_p90_ms": 9968.378499997925, "e2e_p95_ms": 9968.378499997925, "e2e_p99_ms": 9968.378499997925, "e2e_samples": 6, "e2e_stddev_ms": 4313.232538132409, "effective_prefill_tok_s_max": 872008.5168157973, "effective_prefill_tok_s_mean": 475197.9660648341, "effective_prefill_tok_s_mean_ci95_high": 781046.1081230472, "effective_prefill_tok_s_mean_ci95_low": 169349.82400662103, "effective_prefill_tok_s_min": 7363.41904408938, "effective_prefill_tok_s_p25": 7598.657833785861, "effective_prefill_tok_s_p50": 570637.372712217, "effective_prefill_tok_s_p75": 815553.1886446079, "effective_prefill_tok_s_p90": 872008.5168157973, "effective_prefill_tok_s_p95": 872008.5168157973, "effective_prefill_tok_s_p99": 872008.5168157973, "effective_prefill_tok_s_samples": 6, "effective_prefill_tok_s_stddev": 382230.5545004516, "generation_max_ms": 1765.8773000002839, "generation_mean_ci95_high_ms": 1616.749128269172, "generation_mean_ci95_low_ms": 1170.6446050648085, "generation_mean_ms": 1393.6968666669902, "generation_min_ms": 1135.4447000012442, "generation_p25_ms": 1152.5349000003189, "generation_p50_ms": 1153.7046000012197, "generation_p75_ms": 1585.4569999974046, "generation_p90_ms": 1765.8773000002839, "generation_p95_ms": 1765.8773000002839, "generation_p99_ms": 1765.8773000002839, "generation_samples": 6, "generation_stddev_ms": 278.7572586220073, "mean_inter_token_latency_ms_max": 4.663108823521778, "mean_inter_token_latency_ms_mean": 4.290479366822758, "mean_inter_token_latency_ms_mean_ci95_high": 4.59871801399576, "mean_inter_token_latency_ms_mean_ci95_low": 3.982240719649757, "mean_inter_token_latency_ms_min": 3.6509639240544924, "mean_inter_token_latency_ms_p25": 4.107755759166153, "mean_inter_token_latency_ms_p50": 4.204469761905438, "mean_inter_token_latency_ms_p75": 4.628654216868751, "mean_inter_token_latency_ms_p90": 4.663108823521778, "mean_inter_token_latency_ms_p95": 4.663108823521778, "mean_inter_token_latency_ms_p99": 4.663108823521778, "mean_inter_token_latency_ms_samples": 6, "mean_inter_token_latency_ms_stddev": 0.3852180635610356, "mean_time_per_output_token_max_ms": 4.663108823521778, "mean_time_per_output_token_mean_ci95_high_ms": 4.59871801399576, "mean_time_per_output_token_mean_ci95_low_ms": 3.982240719649757, "mean_time_per_output_token_mean_ms": 4.290479366822758, "mean_time_per_output_token_min_ms": 3.6509639240544924, "mean_time_per_output_token_p25_ms": 4.107755759166153, "mean_time_per_output_token_p50_ms": 4.204469761905438, "mean_time_per_output_token_p75_ms": 4.628654216868751, "mean_time_per_output_token_p90_ms": 4.663108823521778, "mean_time_per_output_token_p95_ms": 4.663108823521778, "mean_time_per_output_token_p99_ms": 4.663108823521778, "mean_time_per_output_token_samples": 6, "mean_time_per_output_token_stddev_ms": 0.3852180635610356, "output_token_sources": { "usage": 6 }, "output_tokens": 1966, "prefix_cache_hit_avg": 0.6666131891578702, "prompt_token_samples": 6, "prompt_tokens": 373981, "prompt_tokens_max": 62335.0, "prompt_tokens_mean": 62330.166666666664, "prompt_tokens_mean_ci95_high": 62332.718014894126, "prompt_tokens_mean_ci95_low": 62327.6153184392, "prompt_tokens_min": 62327.0, "prompt_tokens_p25": 62328.0, "prompt_tokens_p50": 62328.0, "prompt_tokens_p75": 62333.0, "prompt_tokens_p90": 62335.0, "prompt_tokens_p95": 62335.0, "prompt_tokens_p99": 62335.0, "prompt_tokens_samples": 6, "prompt_tokens_stddev": 3.1885210782848317, "reasoning_chunks": 417, "request_canary_passed": 6, "request_canary_samples": 6, "throughput_performance_eligible": true, "throughput_tok_s": 77.39950939419339, "time_to_first_output_max_ms": 8464.410299999145, "time_to_first_output_mean_ci95_high_ms": 6245.032571871507, "time_to_first_output_mean_ci95_low_ms": -567.7366718730327, "time_to_first_output_mean_ms": 2838.647949999237, "time_to_first_output_min_ms": 71.48210000013933, "time_to_first_output_p25_ms": 76.42420000047423, "time_to_first_output_p50_ms": 107.8324000009161, "time_to_first_output_p75_ms": 8202.501199997641, "time_to_first_output_p90_ms": 8464.410299999145, "time_to_first_output_p95_ms": 8464.410299999145, "time_to_first_output_p99_ms": 8464.410299999145, "time_to_first_output_samples": 6, "time_to_first_output_stddev_ms": 4257.093975127795, "tpot_max_ms": 4.663108823521778, "tpot_mean_ci95_high_ms": 4.59871801399576, "tpot_mean_ci95_low_ms": 3.982240719649757, "tpot_mean_ms": 4.290479366822758, "tpot_min_ms": 3.6509639240544924, "tpot_p25_ms": 4.107755759166153, "tpot_p50_ms": 4.204469761905438, "tpot_p75_ms": 4.628654216868751, "tpot_p90_ms": 4.663108823521778, "tpot_p95_ms": 4.663108823521778, "tpot_p99_ms": 4.663108823521778, "tpot_samples": 6, "tpot_stddev_ms": 0.3852180635610356, "ttft_max_ms": 9787.702499997977, "ttft_mean_ci95_high_ms": 7536.797318989051, "ttft_mean_ci95_low_ms": 626.4094143424595, "ttft_mean_ms": 4081.6033666657554, "ttft_min_ms": 1049.1322000016226, "ttft_p25_ms": 1079.828300000372, "ttft_p50_ms": 1530.1660999975866, "ttft_p75_ms": 9509.358399998746, "ttft_p90_ms": 9787.702499997977, "ttft_p95_ms": 9787.702499997977, "ttft_p99_ms": 9787.702499997977, "ttft_samples": 6, "ttft_stddev_ms": 4318.092931399175 }, "model": "qwen38-swift15-ninfer-dflash2-c4-262k", "performance_eligible": true, "prompt_cache_mode": "unique", "request_canaries": true, "request_timings": [ { "content_chunks": 7, "controlled_output": null, "decode_tok_s": 273.90026875134754, "e2e_ms": 9618.114900000364, "effective_prefill_tok_s": 7363.41904408938, "generation_ms": 1153.7046000012197, "mean_inter_token_latency_ms": 3.6509639240544924, "mean_time_per_output_token_ms": 3.6509639240544924, "output_token_source": "usage", "output_tokens": 317, "planned_context_tokens": 65536, "prompt_tokens": 62327, "reasoning_chunks": 58, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_1717_00000", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 0, "time_to_first_output_ms": 8464.410299999145, "tpot_ms": 3.6509639240544924, "ttft_ms": 9509.358399998746, "visible_content_truncated": false, "visible_generation_ms": 108.7565000016184 }, { "content_chunks": 10, "controlled_output": null, "decode_tok_s": 222.82018666318382, "e2e_ms": 1211.8689000017184, "effective_prefill_tok_s": 815553.1886446079, "generation_ms": 1135.4447000012442, "mean_inter_token_latency_ms": 4.487923715419938, "mean_time_per_output_token_ms": 4.487923715419938, "output_token_source": "usage", "output_tokens": 254, "planned_context_tokens": 65536, "prompt_tokens": 62328, "reasoning_chunks": 54, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_1717_00001", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 1, "time_to_first_output_ms": 76.42420000047423, "tpot_ms": 4.487923715419938, "ttft_ms": 1049.1322000016226, "visible_content_truncated": false, "visible_generation_ms": 162.73670000009588 }, { "content_chunks": 11, "controlled_output": null, "decode_tok_s": 237.84211960815878, "e2e_ms": 9968.378499997925, "effective_prefill_tok_s": 7598.657833785861, "generation_ms": 1765.8773000002839, "mean_inter_token_latency_ms": 4.204469761905438, "mean_time_per_output_token_ms": 4.204469761905438, "output_token_source": "usage", "output_tokens": 421, "planned_context_tokens": 65536, "prompt_tokens": 62328, "reasoning_chunks": 89, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_1717_00002", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 2, "time_to_first_output_ms": 8202.501199997641, "tpot_ms": 4.204469761905438, "ttft_ms": 9787.702499997977, "visible_content_truncated": false, "visible_generation_ms": 180.67599999994854 }, { "content_chunks": 11, "controlled_output": null, "decode_tok_s": 216.04551844801497, "e2e_ms": 1260.367300001235, "effective_prefill_tok_s": 578026.6413385074, "generation_ms": 1152.5349000003189, "mean_inter_token_latency_ms": 4.628654216868751, "mean_time_per_output_token_ms": 4.628654216868751, "output_token_source": "usage", "output_tokens": 250, "planned_context_tokens": 65536, "prompt_tokens": 62330, "reasoning_chunks": 54, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_1717_00003", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 3, "time_to_first_output_ms": 107.8324000009161, "tpot_ms": 4.628654216868751, "ttft_ms": 1079.828300000372, "visible_content_truncated": false, "visible_generation_ms": 180.53900000086287 }, { "content_chunks": 7, "controlled_output": null, "decode_tok_s": 214.44920928196512, "e2e_ms": 1656.939099997544, "effective_prefill_tok_s": 872008.5168157973, "generation_ms": 1585.4569999974046, "mean_inter_token_latency_ms": 4.663108823521778, "mean_time_per_output_token_ms": 4.663108823521778, "output_token_source": "usage", "output_tokens": 341, "planned_context_tokens": 65536, "prompt_tokens": 62333, "reasoning_chunks": 82, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_1717_00004", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 4, "time_to_first_output_ms": 71.48210000013933, "tpot_ms": 4.663108823521778, "ttft_ms": 1530.1660999975866, "visible_content_truncated": false, "visible_generation_ms": 126.77299999995739 }, { "content_chunks": 9, "controlled_output": null, "decode_tok_s": 243.44193243928243, "e2e_ms": 1678.4001999985776, "effective_prefill_tok_s": 570637.372712217, "generation_ms": 1569.1627000014705, "mean_inter_token_latency_ms": 4.107755759166153, "mean_time_per_output_token_ms": 4.107755759166153, "output_token_source": "usage", "output_tokens": 383, "planned_context_tokens": 65536, "prompt_tokens": 62335, "reasoning_chunks": 80, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_1717_00005", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 5, "time_to_first_output_ms": 109.23749999710708, "tpot_ms": 4.107755759166153, "ttft_ms": 1533.4326999982295, "visible_content_truncated": false, "visible_generation_ms": 144.96750000034808 } ], "requests": 6, "response_words": 0, "responses_received": 6, "run_id": "benchmark-20261005T114404Z", "sampling": { "temperature": { "effective_request": 0.9, "requested": 0.9, "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": null, "shared_prefix_burst": false, "thinking_mode": "default" }, "shared_prefix_tokens": 0, "source_recipe": null, "started_at": "2026-10-05T11:44:04Z", "started_at_unix_ns": 1791200644799903400, "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": 25400.677800000267 }