{ "base_url": "http://127.0.0.1:39273/v1", "completed": 5, "concurrency": 1, "context_distribution": { "4096": 5 }, "context_seed": 0, "context_tokens": 4096, "engine": "sglang", "failed": 0, "failures": [], "finished_at": "2026-09-05T03:28:20Z", "gpu": "Primary Node / RTX PRO 6000 Blackwell Max-Q / TP=1", "manifest": null, "max_context_tokens": 262144, "max_tokens": 512, "measurement_protocol": "capacity-v3", "metrics": { "content_chunks": 67, "content_chunks_s": null, "decode_tok_s_max": 131.89318245286825, "decode_tok_s_mean": 115.72989869004604, "decode_tok_s_min": 96.00877234787538, "decode_tok_s_p25": 111.67028975427336, "decode_tok_s_p50": 119.49925234982315, "decode_tok_s_p75": 119.57799654539016, "decode_tok_s_p90": 131.89318245286825, "decode_tok_s_p95": 131.89318245286825, "decode_tok_s_p99": 131.89318245286825, "decode_tok_s_samples": 5, "e2e_max_ms": 838.9871999388561, "e2e_mean_ms": 573.8655799767002, "e2e_min_ms": 430.95850001554936, "e2e_p25_ms": 499.0016999654472, "e2e_p50_ms": 505.10720000602305, "e2e_p75_ms": 595.2732999576256, "e2e_p90_ms": 838.9871999388561, "e2e_p95_ms": 838.9871999388561, "e2e_p99_ms": 838.9871999388561, "e2e_samples": 5, "effective_prefill_tok_s_max": 34430.14911884262, "effective_prefill_tok_s_mean": 28769.600048887216, "effective_prefill_tok_s_min": 8830.493628501343, "effective_prefill_tok_s_p25": 32316.839207606383, "effective_prefill_tok_s_p50": 34100.08984530743, "effective_prefill_tok_s_p75": 34170.42844417832, "effective_prefill_tok_s_p90": 34430.14911884262, "effective_prefill_tok_s_p95": 34430.14911884262, "effective_prefill_tok_s_p99": 34430.14911884262, "effective_prefill_tok_s_samples": 5, "generation_max_ms": 489.5385999698192, "generation_mean_ms": 406.35071999859065, "generation_min_ms": 326.02140004746616, "generation_p25_ms": 393.0488999467343, "generation_p50_ms": 393.30790005624294, "generation_p75_ms": 429.83679997269064, "generation_p90_ms": 489.5385999698192, "generation_p95_ms": 489.5385999698192, "generation_p99_ms": 489.5385999698192, "generation_samples": 5, "mean_inter_token_latency_ms_max": 10.415714892974876, "mean_inter_token_latency_ms_mean": 8.736707398968628, "mean_inter_token_latency_ms_min": 7.581893024359678, "mean_inter_token_latency_ms_p25": 8.362742552058176, "mean_inter_token_latency_ms_p50": 8.36825319268602, "mean_inter_token_latency_ms_p75": 8.954933332764389, "mean_inter_token_latency_ms_p90": 10.415714892974876, "mean_inter_token_latency_ms_p95": 10.415714892974876, "mean_inter_token_latency_ms_p99": 10.415714892974876, "mean_inter_token_latency_ms_samples": 5, "output_token_sources": { "usage": 5 }, "output_tokens": 237, "prefix_cache_hit_avg": null, "prompt_token_samples": 5, "prompt_tokens": 18065, "prompt_tokens_max": 3613.0, "prompt_tokens_mean": 3613.0, "prompt_tokens_min": 3613.0, "prompt_tokens_p25": 3613.0, "prompt_tokens_p50": 3613.0, "prompt_tokens_p75": 3613.0, "prompt_tokens_p90": 3613.0, "prompt_tokens_p95": 3613.0, "prompt_tokens_p99": 3613.0, "prompt_tokens_samples": 5, "reasoning_chunks": 0, "throughput_tok_s": 82.56458849968638, "time_to_first_output_max_ms": 409.1503999661654, "time_to_first_output_mean_ms": 167.5148599781096, "time_to_first_output_min_ms": 104.9370999680832, "time_to_first_output_p25_ms": 105.73469998780638, "time_to_first_output_p50_ms": 105.95280001871288, "time_to_first_output_p75_ms": 111.7992999497801, "time_to_first_output_p90_ms": 409.1503999661654, "time_to_first_output_p95_ms": 409.1503999661654, "time_to_first_output_p99_ms": 409.1503999661654, "time_to_first_output_samples": 5, "tpot_max_ms": 10.415714892974876, "tpot_mean_ms": 8.736707398968628, "tpot_min_ms": 7.581893024359678, "tpot_p25_ms": 8.362742552058176, "tpot_p50_ms": 8.36825319268602, "tpot_p75_ms": 8.954933332764389, "tpot_p90_ms": 10.415714892974876, "tpot_p95_ms": 10.415714892974876, "tpot_p99_ms": 10.415714892974876, "tpot_samples": 5, "ttft_max_ms": 409.15129997301847, "ttft_mean_ms": 167.51529998145998, "ttft_min_ms": 104.93739997036755, "ttft_p25_ms": 105.73499999009073, "ttft_p50_ms": 105.95300002023578, "ttft_p75_ms": 111.79979995358735, "ttft_p90_ms": 409.15129997301847, "ttft_p95_ms": 409.15129997301847, "ttft_p99_ms": 409.15129997301847, "ttft_samples": 5 }, "model": "qwen38-27b-inferact-nvfp4-pro6000-dflash2-k12-chunk8k", "request_timings": [ { "content_chunks": 14, "decode_tok_s": 111.67028975427336, "e2e_ms": 838.9871999388561, "effective_prefill_tok_s": 8830.493628501343, "generation_ms": 429.83679997269064, "mean_inter_token_latency_ms": 8.954933332764389, "output_token_source": "usage", "output_tokens": 49, "planned_context_tokens": 4096, "prompt_tokens": 3613, "reasoning_chunks": 0, "request_index": 0, "time_to_first_output_ms": 409.1503999661654, "tpot_ms": 8.954933332764389, "ttft_ms": 409.15129997301847, "visible_generation_ms": 429.8358999658376 }, { "content_chunks": 13, "decode_tok_s": 119.49925234982315, "e2e_ms": 505.10720000602305, "effective_prefill_tok_s": 32316.839207606383, "generation_ms": 393.30790005624294, "mean_inter_token_latency_ms": 8.36825319268602, "output_token_source": "usage", "output_tokens": 48, "planned_context_tokens": 4096, "prompt_tokens": 3613, "reasoning_chunks": 0, "request_index": 1, "time_to_first_output_ms": 111.7992999497801, "tpot_ms": 8.36825319268602, "ttft_ms": 111.79979995358735, "visible_generation_ms": 393.3074000524357 }, { "content_chunks": 16, "decode_tok_s": 96.00877234787538, "e2e_ms": 595.2732999576256, "effective_prefill_tok_s": 34170.42844417832, "generation_ms": 489.5385999698192, "mean_inter_token_latency_ms": 10.415714892974876, "output_token_source": "usage", "output_tokens": 48, "planned_context_tokens": 4096, "prompt_tokens": 3613, "reasoning_chunks": 0, "request_index": 2, "time_to_first_output_ms": 105.73469998780638, "tpot_ms": 10.415714892974876, "ttft_ms": 105.73499999009073, "visible_generation_ms": 489.53829996753484 }, { "content_chunks": 11, "decode_tok_s": 131.89318245286825, "e2e_ms": 430.95850001554936, "effective_prefill_tok_s": 34430.14911884262, "generation_ms": 326.02140004746616, "mean_inter_token_latency_ms": 7.581893024359678, "output_token_source": "usage", "output_tokens": 44, "planned_context_tokens": 4096, "prompt_tokens": 3613, "reasoning_chunks": 0, "request_index": 3, "time_to_first_output_ms": 104.9370999680832, "tpot_ms": 7.581893024359678, "ttft_ms": 104.93739997036755, "visible_generation_ms": 326.0211000451818 }, { "content_chunks": 13, "decode_tok_s": 119.57799654539016, "e2e_ms": 499.0016999654472, "effective_prefill_tok_s": 34100.08984530743, "generation_ms": 393.0488999467343, "mean_inter_token_latency_ms": 8.362742552058176, "output_token_source": "usage", "output_tokens": 48, "planned_context_tokens": 4096, "prompt_tokens": 3613, "reasoning_chunks": 0, "request_index": 4, "time_to_first_output_ms": 105.95280001871288, "tpot_ms": 8.362742552058176, "ttft_ms": 105.95300002023578, "visible_generation_ms": 393.0486999452114 } ], "requests": 5, "run_id": "benchmark-20260905T032817Z", "schema": "anvil-serving.benchmark/v1", "serve_flags": { "no_thinking": false, "reasoning_effort": null, "shared_prefix_burst": false, "thinking_mode": "disabled" }, "source_recipe": null, "started_at": "2026-09-05T03:28:17Z", "tier": null, "timing_methodology": { "clock": "client time.perf_counter", "decode": "usage completion tokens after the first token divided by client-observed generation time; completion tokens may include reasoning tokens", "distributions": "request-level 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": "client-observed generation time divided by completion tokens after the first token; not a raw per-token timestamp trace", "time_to_first_output": "request start through the first non-empty streamed reasoning or content delta", "tpot": "alias of mean_inter_token_latency for capacity-v3: client-observed generation time divided by completion tokens after the first token", "ttft": "request start through the first non-empty streamed content delta" }, "wall_clock_ms": 2870.4800000414252 }