{ "base_url": "http://127.0.0.1:30000/v1", "clock_domain_id": null, "completed": 2, "concurrency": 4, "configuration_fingerprint": null, "context_distribution": { "8192": 4 }, "context_seed": 0, "context_tokens": 8192, "controlled_output_policy": "strict", "engine": "vllm", "failed": 2, "failures": [ { "controlled_output": { "capture_complete": true, "exact_adherence": false, "observed_code_words": 63, "observed_extra_words": 0, "passed": false, "policy": "strict", "requested_words": 64 }, "elapsed_ms": 17405.23385896813, "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": 72, "finish_reasons": [ "stop" ], "output_token_source": "usage", "reasoning_chunks": 269, "stream_terminal_observed": true, "usage": { "completion_tokens": 359, "prompt_tokens": 6066, "total_tokens": 6425 }, "visible_content_truncated": false }, "response_received": true, "response_timing": { "content_chunks": 72, "controlled_output": { "capture_complete": true, "exact_adherence": false, "observed_code_words": 63, "observed_extra_words": 0, "passed": false, "policy": "strict", "requested_words": 64 }, "decode_tok_s": 22.41586518412521, "e2e_ms": 17405.06630798336, "effective_prefill_tok_s": 4229.432220730992, "generation_ms": 15970.831242040731, "mean_inter_token_latency_ms": 44.61126045262774, "mean_time_per_output_token_ms": 44.61126045262774, "output_token_source": "usage", "output_tokens": 359, "planned_context_tokens": 8192, "prompt_tokens": 6066, "reasoning_chunks": 269, "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": 1434.2350659426302, "tpot_ms": 44.61126045262774, "ttft_ms": 16050.450776005164, "visible_content_truncated": false, "visible_generation_ms": 1354.6155319781974 } }, { "controlled_output": { "capture_complete": true, "exact_adherence": false, "observed_code_words": 63, "observed_extra_words": 0, "passed": false, "policy": "strict", "requested_words": 64 }, "elapsed_ms": 17168.46519196406, "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_00001", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 1, "response_metadata": { "content_chunks": 72, "finish_reasons": [ "stop" ], "output_token_source": "usage", "reasoning_chunks": 255, "stream_terminal_observed": true, "usage": { "completion_tokens": 343, "prompt_tokens": 6068, "total_tokens": 6411 }, "visible_content_truncated": false }, "response_received": true, "response_timing": { "content_chunks": 72, "controlled_output": { "capture_complete": true, "exact_adherence": false, "observed_code_words": 63, "observed_extra_words": 0, "passed": false, "policy": "strict", "requested_words": 64 }, "decode_tok_s": 23.872596132520496, "e2e_ms": 17168.101551011205, "effective_prefill_tok_s": 2135.0772039787425, "generation_ms": 14326.049756025895, "mean_inter_token_latency_ms": 41.889034374344725, "mean_time_per_output_token_ms": 41.889034374344725, "output_token_source": "usage", "output_tokens": 343, "planned_context_tokens": 8192, "prompt_tokens": 6068, "reasoning_chunks": 255, "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": 2842.0517949853092, "tpot_ms": 41.889034374344725, "ttft_ms": 15802.488909917884, "visible_content_truncated": false, "visible_generation_ms": 1365.6126410933211 } } ], "finished_at": "2026-09-17T17:15:29Z", "finished_at_unix_ns": 1789665329213286968, "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": 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": 2, "e2e": 2, "effective_prefill_tok_s": 2, "tpot": 2, "ttft": 2 }, "performance_eligible_requests": 2, "successful_requests": 2 }, "metrics": { "completion_tokens_max": 663.0, "completion_tokens_mean": 659.0, "completion_tokens_mean_ci95_high": 666.84, "completion_tokens_mean_ci95_low": 651.16, "completion_tokens_min": 655.0, "completion_tokens_p25": 655.0, "completion_tokens_p50": 655.0, "completion_tokens_p75": 663.0, "completion_tokens_p90": 663.0, "completion_tokens_p95": 663.0, "completion_tokens_p99": 663.0, "completion_tokens_samples": 2, "completion_tokens_stddev": 5.656854249492381, "content_chunks": 146, "content_chunks_s": null, "controlled_output_exact_adherence": 2, "controlled_output_exact_adherence_rate": 1.0, "controlled_output_samples": 2, "controlled_output_unobservable": 0, "decode_tok_s_max": 58.782862368247045, "decode_tok_s_mean": 47.89182840533688, "decode_tok_s_mean_ci95_high": 69.23825497264082, "decode_tok_s_mean_ci95_low": 26.54540183803295, "decode_tok_s_min": 37.00079444242671, "decode_tok_s_p25": 37.00079444242671, "decode_tok_s_p50": 37.00079444242671, "decode_tok_s_p75": 58.782862368247045, "decode_tok_s_p90": 58.782862368247045, "decode_tok_s_p95": 58.782862368247045, "decode_tok_s_p99": 58.782862368247045, "decode_tok_s_samples": 2, "decode_tok_s_stddev": 15.402247938613556, "e2e_max_ms": 22077.97193794977, "e2e_mean_ci95_high_ms": 22149.248611512594, "e2e_mean_ci95_low_ms": 21858.202194464393, "e2e_mean_ms": 22003.725402988493, "e2e_min_ms": 21929.478868027218, "e2e_p25_ms": 21929.478868027218, "e2e_p50_ms": 21929.478868027218, "e2e_p75_ms": 22077.97193794977, "e2e_p90_ms": 22077.97193794977, "e2e_p95_ms": 22077.97193794977, "e2e_p99_ms": 22077.97193794977, "e2e_samples": 2, "e2e_stddev_ms": 105.00045670144434, "effective_prefill_tok_s_max": 1426.3609297820678, "effective_prefill_tok_s_mean": 993.7322175010977, "effective_prefill_tok_s_mean_ci95_high": 1841.684493571799, "effective_prefill_tok_s_mean_ci95_low": 145.77994143039643, "effective_prefill_tok_s_min": 561.1035052201275, "effective_prefill_tok_s_p25": 561.1035052201275, "effective_prefill_tok_s_p50": 561.1035052201275, "effective_prefill_tok_s_p75": 1426.3609297820678, "effective_prefill_tok_s_p90": 1426.3609297820678, "effective_prefill_tok_s_p95": 1426.3609297820678, "effective_prefill_tok_s_p99": 1426.3609297820678, "effective_prefill_tok_s_samples": 2, "effective_prefill_tok_s_stddev": 611.8293923797555, "generation_max_ms": 17675.296162022278, "generation_mean_ci95_high_ms": 20753.78117418848, "generation_mean_ci95_low_ms": 8183.300707843156, "generation_mean_ms": 14468.540941015817, "generation_min_ms": 11261.785720009357, "generation_p25_ms": 11261.785720009357, "generation_p50_ms": 11261.785720009357, "generation_p75_ms": 17675.296162022278, "generation_p90_ms": 17675.296162022278, "generation_p95_ms": 17675.296162022278, "generation_p99_ms": 17675.296162022278, "generation_samples": 2, "generation_stddev_ms": 4535.036724758068, "mean_inter_token_latency_ms_max": 27.026446730920913, "mean_inter_token_latency_ms_mean": 22.019103818639728, "mean_inter_token_latency_ms_mean_ci95_high": 31.833495926710846, "mean_inter_token_latency_ms_mean_ci95_low": 12.20471171056861, "mean_inter_token_latency_ms_min": 17.011760906358546, "mean_inter_token_latency_ms_p25": 17.011760906358546, "mean_inter_token_latency_ms_p50": 17.011760906358546, "mean_inter_token_latency_ms_p75": 27.026446730920913, "mean_inter_token_latency_ms_p90": 27.026446730920913, "mean_inter_token_latency_ms_p95": 27.026446730920913, "mean_inter_token_latency_ms_p99": 27.026446730920913, "mean_inter_token_latency_ms_samples": 2, "mean_inter_token_latency_ms_stddev": 7.081452258000841, "mean_time_per_output_token_max_ms": 27.026446730920913, "mean_time_per_output_token_mean_ci95_high_ms": 31.833495926710846, "mean_time_per_output_token_mean_ci95_low_ms": 12.20471171056861, "mean_time_per_output_token_mean_ms": 22.019103818639728, "mean_time_per_output_token_min_ms": 17.011760906358546, "mean_time_per_output_token_p25_ms": 17.011760906358546, "mean_time_per_output_token_p50_ms": 17.011760906358546, "mean_time_per_output_token_p75_ms": 27.026446730920913, "mean_time_per_output_token_p90_ms": 27.026446730920913, "mean_time_per_output_token_p95_ms": 27.026446730920913, "mean_time_per_output_token_p99_ms": 27.026446730920913, "mean_time_per_output_token_samples": 2, "mean_time_per_output_token_stddev_ms": 7.081452258000841, "output_token_sources": { "usage": 2 }, "output_tokens": 1318, "prefix_cache_hit_avg": null, "prompt_token_samples": 2, "prompt_tokens": 12137, "prompt_tokens_max": 6069.0, "prompt_tokens_mean": 6068.5, "prompt_tokens_mean_ci95_high": 6069.48, "prompt_tokens_mean_ci95_low": 6067.52, "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": 2, "prompt_tokens_stddev": 0.7071067811865476, "reasoning_chunks": 1009, "request_canary_passed": 2, "request_canary_samples": 2, "throughput_performance_eligible": false, "throughput_tok_s": null, "time_to_first_output_max_ms": 10816.186217940412, "time_to_first_output_mean_ci95_high_ms": 13965.94790366944, "time_to_first_output_mean_ci95_low_ms": 1104.4210202759132, "time_to_first_output_mean_ms": 7535.184461972676, "time_to_first_output_min_ms": 4254.18270600494, "time_to_first_output_p25_ms": 4254.18270600494, "time_to_first_output_p50_ms": 4254.18270600494, "time_to_first_output_p75_ms": 10816.186217940412, "time_to_first_output_p90_ms": 10816.186217940412, "time_to_first_output_p95_ms": 10816.186217940412, "time_to_first_output_p99_ms": 10816.186217940412, "time_to_first_output_samples": 2, "time_to_first_output_stddev_ms": 4640.037181459513, "tpot_max_ms": 27.026446730920913, "tpot_mean_ci95_high_ms": 31.833495926710846, "tpot_mean_ci95_low_ms": 12.20471171056861, "tpot_mean_ms": 22.019103818639728, "tpot_min_ms": 17.011760906358546, "tpot_p25_ms": 17.011760906358546, "tpot_p50_ms": 17.011760906358546, "tpot_p75_ms": 27.026446730920913, "tpot_p90_ms": 27.026446730920913, "tpot_p95_ms": 27.026446730920913, "tpot_p99_ms": 27.026446730920913, "tpot_samples": 2, "tpot_stddev_ms": 7.081452258000841, "ttft_max_ms": 21014.167870976962, "ttft_mean_ci95_high_ms": 21093.894126662053, "ttft_mean_ci95_low_ms": 20768.345249281265, "ttft_mean_ms": 20931.11968797166, "ttft_min_ms": 20848.071504966356, "ttft_p25_ms": 20848.071504966356, "ttft_p50_ms": 20848.071504966356, "ttft_p75_ms": 21014.167870976962, "ttft_p90_ms": 21014.167870976962, "ttft_p95_ms": 21014.167870976962, "ttft_p99_ms": 21014.167870976962, "ttft_samples": 2, "ttft_stddev_ms": 117.4478667365425 }, "model": "glm53-flash-exl3-4bpw-v84-fp8-tp2-c4-327k-mtp3-r7-nospec", "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": 37.00079444242671, "e2e_ms": 21929.478868027218, "effective_prefill_tok_s": 1426.3609297820678, "generation_ms": 17675.296162022278, "mean_inter_token_latency_ms": 27.026446730920913, "mean_time_per_output_token_ms": 27.026446730920913, "output_token_source": "usage", "output_tokens": 655, "planned_context_tokens": 8192, "prompt_tokens": 6068, "reasoning_chunks": 500, "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": 4254.18270600494, "tpot_ms": 27.026446730920913, "ttft_ms": 20848.071504966356, "visible_content_truncated": false, "visible_generation_ms": 1081.4073630608618 }, { "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": 58.782862368247045, "e2e_ms": 22077.97193794977, "effective_prefill_tok_s": 561.1035052201275, "generation_ms": 11261.785720009357, "mean_inter_token_latency_ms": 17.011760906358546, "mean_time_per_output_token_ms": 17.011760906358546, "output_token_source": "usage", "output_tokens": 663, "planned_context_tokens": 8192, "prompt_tokens": 6069, "reasoning_chunks": 509, "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": 10816.186217940412, "tpot_ms": 17.011760906358546, "ttft_ms": 21014.167870976962, "visible_content_truncated": false, "visible_generation_ms": 1063.804066972807 } ], "requests": 4, "response_words": 64, "responses_received": 4, "run_id": "benchmark-20260917T171507Z", "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": "2026-09-17T17:15:07Z", "started_at_unix_ns": 1789665307133539886, "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": 22079.744492075406 }