{ "base_url": "http://127.0.0.1:30002/v1", "clock_domain_id": null, "completed": 3, "concurrency": 4, "configuration_fingerprint": null, "context_distribution": { "8192": 4 }, "context_seed": 0, "context_tokens": 8192, "controlled_output_policy": "strict", "engine": "vllm", "failed": 1, "failures": [ { "controlled_output": { "capture_complete": true, "exact_adherence": false, "observed_code_words": 65, "observed_extra_words": 0, "passed": false, "policy": "strict", "requested_words": 64 }, "elapsed_ms": 12408.61093904823, "error_message": "controlled output validation failed", "error_message_truncated": false, "error_type": "ValueError", "failure_class": "controlled_output_failed", "performance_eligible": false, "planned_context_tokens": 8192, "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, "response_metadata": { "content_chunks": 74, "finish_reasons": [ "stop" ], "output_token_source": "usage", "reasoning_chunks": 256, "stream_terminal_observed": true, "usage": { "completion_tokens": 346, "prompt_tokens": 6066, "total_tokens": 6412 }, "visible_content_truncated": false }, "response_received": true, "response_timing": { "content_chunks": 74, "controlled_output": { "capture_complete": true, "exact_adherence": false, "observed_code_words": 65, "observed_extra_words": 0, "passed": false, "policy": "strict", "requested_words": 64 }, "decode_tok_s": 37.91248387655915, "e2e_ms": 12408.208559034392, "effective_prefill_tok_s": 1833.5681157500057, "generation_ms": 9099.904958042316, "mean_inter_token_latency_ms": 26.376536110267583, "mean_time_per_output_token_ms": 26.376536110267583, "output_token_source": "usage", "output_tokens": 346, "planned_context_tokens": 8192, "prompt_tokens": 6066, "reasoning_chunks": 256, "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": 3308.303600992076, "tpot_ms": 26.376536110267583, "ttft_ms": 11117.213090066798, "visible_content_truncated": false, "visible_generation_ms": 1290.9954689675942 } } ], "finished_at": "2026-09-17T17:21:51Z", "finished_at_unix_ns": 1789665711506672407, "gpu": "dual-sm120", "manifest": null, "max_context_tokens": 327680, "max_tokens": 16384, "measurement_protocol": "capacity-v4-reasoning", "metric_population": { "aggregate_throughput_eligible": false, "attempted_requests": 4, "excluded_requests": 1, "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": 1018.0, "completion_tokens_mean": 707.3333333333334, "completion_tokens_mean_ci95_high": 1044.4862516228109, "completion_tokens_mean_ci95_low": 370.1804150438558, "completion_tokens_min": 424.0, "completion_tokens_p25": 424.0, "completion_tokens_p50": 680.0, "completion_tokens_p75": 1018.0, "completion_tokens_p90": 1018.0, "completion_tokens_p95": 1018.0, "completion_tokens_p99": 1018.0, "completion_tokens_samples": 3, "completion_tokens_stddev": 297.9418287742312, "content_chunks": 219, "content_chunks_s": null, "controlled_output_exact_adherence": 3, "controlled_output_exact_adherence_rate": 1.0, "controlled_output_samples": 3, "controlled_output_unobservable": 0, "decode_tok_s_max": 62.745064746661626, "decode_tok_s_mean": 53.29212958513181, "decode_tok_s_mean_ci95_high": 70.72446551101211, "decode_tok_s_mean_ci95_low": 35.85979365925151, "decode_tok_s_min": 35.515994102077684, "decode_tok_s_p25": 35.515994102077684, "decode_tok_s_p50": 61.615329906656115, "decode_tok_s_p75": 62.745064746661626, "decode_tok_s_p90": 62.745064746661626, "decode_tok_s_p95": 62.745064746661626, "decode_tok_s_p99": 62.745064746661626, "decode_tok_s_samples": 3, "decode_tok_s_stddev": 15.404944652159658, "e2e_max_ms": 21465.36522707902, "e2e_mean_ci95_high_ms": 21940.88299349537, "e2e_mean_ci95_low_ms": 13004.1519891412, "e2e_mean_ms": 17472.517491318285, "e2e_min_ms": 13569.514815928414, "e2e_p25_ms": 13569.514815928414, "e2e_p50_ms": 17382.672430947423, "e2e_p75_ms": 21465.36522707902, "e2e_p90_ms": 21465.36522707902, "e2e_p95_ms": 21465.36522707902, "e2e_p99_ms": 21465.36522707902, "e2e_samples": 3, "e2e_stddev_ms": 3948.691875795271, "effective_prefill_tok_s_max": 3656.770233096792, "effective_prefill_tok_s_mean": 1935.0733438168845, "effective_prefill_tok_s_mean_ci95_high": 3630.7656859418585, "effective_prefill_tok_s_mean_ci95_low": 239.38100169191034, "effective_prefill_tok_s_min": 924.9967421877255, "effective_prefill_tok_s_p25": 924.9967421877255, "effective_prefill_tok_s_p50": 1223.4530561661359, "effective_prefill_tok_s_p75": 3656.770233096792, "effective_prefill_tok_s_p90": 3656.770233096792, "effective_prefill_tok_s_p95": 3656.770233096792, "effective_prefill_tok_s_p99": 3656.770233096792, "effective_prefill_tok_s_samples": 3, "effective_prefill_tok_s_stddev": 1498.4822911050624, "generation_max_ms": 16505.632632994093, "generation_mean_ci95_high_ms": 16493.11886809083, "generation_mean_ci95_low_ms": 9665.099709801378, "generation_mean_ms": 13079.109288946105, "generation_min_ms": 10821.568241924979, "generation_p25_ms": 10821.568241924979, "generation_p50_ms": 11910.126991919242, "generation_p75_ms": 16505.632632994093, "generation_p90_ms": 16505.632632994093, "generation_p95_ms": 16505.632632994093, "generation_p99_ms": 16505.632632994093, "generation_samples": 3, "generation_stddev_ms": 3016.958188064035, "mean_inter_token_latency_ms_max": 28.156328586097498, "mean_inter_token_latency_ms_mean": 20.107854770651606, "mean_inter_token_latency_ms_mean_ci95_high": 27.99709183595595, "mean_inter_token_latency_ms_mean_ci95_low": 12.218617705347265, "mean_inter_token_latency_ms_min": 15.937508456443268, "mean_inter_token_latency_ms_p25": 15.937508456443268, "mean_inter_token_latency_ms_p50": 16.229727269414052, "mean_inter_token_latency_ms_p75": 28.156328586097498, "mean_inter_token_latency_ms_p90": 28.156328586097498, "mean_inter_token_latency_ms_p95": 28.156328586097498, "mean_inter_token_latency_ms_p99": 28.156328586097498, "mean_inter_token_latency_ms_samples": 3, "mean_inter_token_latency_ms_stddev": 6.97171399492995, "mean_time_per_output_token_max_ms": 28.156328586097498, "mean_time_per_output_token_mean_ci95_high_ms": 27.99709183595595, "mean_time_per_output_token_mean_ci95_low_ms": 12.218617705347265, "mean_time_per_output_token_mean_ms": 20.107854770651606, "mean_time_per_output_token_min_ms": 15.937508456443268, "mean_time_per_output_token_p25_ms": 15.937508456443268, "mean_time_per_output_token_p50_ms": 16.229727269414052, "mean_time_per_output_token_p75_ms": 28.156328586097498, "mean_time_per_output_token_p90_ms": 28.156328586097498, "mean_time_per_output_token_p95_ms": 28.156328586097498, "mean_time_per_output_token_p99_ms": 28.156328586097498, "mean_time_per_output_token_samples": 3, "mean_time_per_output_token_stddev_ms": 6.97171399492995, "output_token_sources": { "usage": 3 }, "output_tokens": 2122, "prefix_cache_hit_avg": null, "prompt_token_samples": 3, "prompt_tokens": 18205, "prompt_tokens_max": 6069.0, "prompt_tokens_mean": 6068.333333333333, "prompt_tokens_mean_ci95_high": 6068.986666666667, "prompt_tokens_mean_ci95_low": 6067.679999999999, "prompt_tokens_min": 6068.0, "prompt_tokens_p25": 6068.0, "prompt_tokens_p50": 6068.0, "prompt_tokens_p75": 6069.0, "prompt_tokens_p90": 6069.0, "prompt_tokens_p95": 6069.0, "prompt_tokens_p99": 6069.0, "prompt_tokens_samples": 3, "prompt_tokens_stddev": 0.5773502691896257, "reasoning_chunks": 1718, "request_canary_passed": 3, "request_canary_samples": 3, "throughput_performance_eligible": false, "throughput_tok_s": null, "time_to_first_output_max_ms": 6561.104189022444, "time_to_first_output_mean_ci95_high_ms": 7221.801688633427, "time_to_first_output_mean_ci95_low_ms": 1565.0147161109353, "time_to_first_output_mean_ms": 4393.408202372181, "time_to_first_output_min_ms": 1659.3878240091726, "time_to_first_output_p25_ms": 1659.3878240091726, "time_to_first_output_p50_ms": 4959.732594084926, "time_to_first_output_p75_ms": 6561.104189022444, "time_to_first_output_p90_ms": 6561.104189022444, "time_to_first_output_p95_ms": 6561.104189022444, "time_to_first_output_p99_ms": 6561.104189022444, "time_to_first_output_samples": 3, "time_to_first_output_stddev_ms": 2499.44960306191, "tpot_max_ms": 28.156328586097498, "tpot_mean_ci95_high_ms": 27.99709183595595, "tpot_mean_ci95_low_ms": 12.218617705347265, "tpot_mean_ms": 20.107854770651606, "tpot_min_ms": 15.937508456443268, "tpot_p25_ms": 15.937508456443268, "tpot_p50_ms": 16.229727269414052, "tpot_p75_ms": 28.156328586097498, "tpot_p90_ms": 28.156328586097498, "tpot_p95_ms": 28.156328586097498, "tpot_p99_ms": 28.156328586097498, "tpot_samples": 3, "tpot_stddev_ms": 6.97171399492995, "ttft_max_ms": 20576.466378988698, "ttft_mean_ci95_high_ms": 21054.151458686923, "ttft_mean_ci95_low_ms": 11845.178926575223, "ttft_mean_ms": 16449.665192631073, "ttft_min_ms": 12441.079222946428, "ttft_p25_ms": 12441.079222946428, "ttft_p50_ms": 16331.449975958094, "ttft_p75_ms": 20576.466378988698, "ttft_p90_ms": 20576.466378988698, "ttft_p95_ms": 20576.466378988698, "ttft_p99_ms": 20576.466378988698, "ttft_samples": 3, "ttft_stddev_ms": 4068.981712021346 }, "model": "glm53-mixed35-v84-stream-r7-c4-v2", "performance_eligible": false, "prompt_cache_mode": "unique", "request_canaries": true, "request_timings": [ { "content_chunks": 73, "controlled_output": { "capture_complete": true, "exact_adherence": true, "observed_code_words": 64, "observed_extra_words": 0, "passed": true, "policy": "strict", "requested_words": 64 }, "decode_tok_s": 35.515994102077684, "e2e_ms": 13569.514815928414, "effective_prefill_tok_s": 3656.770233096792, "generation_ms": 11910.126991919242, "mean_inter_token_latency_ms": 28.156328586097498, "mean_time_per_output_token_ms": 28.156328586097498, "output_token_source": "usage", "output_tokens": 424, "planned_context_tokens": 8192, "prompt_tokens": 6068, "reasoning_chunks": 336, "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": 1659.3878240091726, "tpot_ms": 28.156328586097498, "ttft_ms": 12441.079222946428, "visible_content_truncated": false, "visible_generation_ms": 1128.4355929819867 }, { "content_chunks": 73, "controlled_output": { "capture_complete": true, "exact_adherence": true, "observed_code_words": 64, "observed_extra_words": 0, "passed": true, "policy": "strict", "requested_words": 64 }, "decode_tok_s": 61.615329906656115, "e2e_ms": 21465.36522707902, "effective_prefill_tok_s": 1223.4530561661359, "generation_ms": 16505.632632994093, "mean_inter_token_latency_ms": 16.229727269414052, "mean_time_per_output_token_ms": 16.229727269414052, "output_token_source": "usage", "output_tokens": 1018, "planned_context_tokens": 8192, "prompt_tokens": 6068, "reasoning_chunks": 858, "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": 4959.732594084926, "tpot_ms": 16.229727269414052, "ttft_ms": 20576.466378988698, "visible_content_truncated": false, "visible_generation_ms": 888.8988480903208 }, { "content_chunks": 73, "controlled_output": { "capture_complete": true, "exact_adherence": true, "observed_code_words": 64, "observed_extra_words": 0, "passed": true, "policy": "strict", "requested_words": 64 }, "decode_tok_s": 62.745064746661626, "e2e_ms": 17382.672430947423, "effective_prefill_tok_s": 924.9967421877255, "generation_ms": 10821.568241924979, "mean_inter_token_latency_ms": 15.937508456443268, "mean_time_per_output_token_ms": 15.937508456443268, "output_token_source": "usage", "output_tokens": 680, "planned_context_tokens": 8192, "prompt_tokens": 6069, "reasoning_chunks": 524, "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": 6561.104189022444, "tpot_ms": 15.937508456443268, "ttft_ms": 16331.449975958094, "visible_content_truncated": false, "visible_generation_ms": 1051.222454989329 } ], "requests": 4, "response_words": 64, "responses_received": 4, "run_id": "benchmark-20260917T172130Z", "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": "max", "shared_prefix_burst": false, "thinking_mode": "default" }, "shared_prefix_tokens": 0, "source_recipe": null, "started_at": "2026-09-17T17:21:30Z", "started_at_unix_ns": 1789665690040003726, "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": 21466.666001011617 }