{ "provenance": "Re-extracted from existing 2026-10-13/14 recordings; no new benchmark.", "selected_requests": [ { "group": "confirm-candidate-off", "id": "r1--long_read", "case_id": "long_read", "complete": true, "finish_reason": "stop", "ttft_s": 3.43465989199467, "e2e_s": 4.1846920079988195, "input_tokens": 4392, "output_tokens": 67, "reused_tokens": 0, "prefill_tps": 1287.1, "decode_tps": 87.8, "decode_s": 0.7632000000000001, "draft_accepted": 51, "draft_offered": 55, "acceptance": 0.9272727272727272, "tail_skipped_experts": 0, "read_mb": 0.18432, "major_faults": 2, "cpu_core_equiv": 3.8115110907833625, "output_sha256": "97e20fe77c97f95265f9b7ce825fac930f247754d7e92b2c6d580ee16e507184", "gpu_util_mean": 89.0, "gpu_power_mean_w": 291.985, "gpu_temperature_max": 72.0, "gpu_vram_peak_mib": 24093.0, "ram_available_min_gib": 11.480022430419922, "bare_json": false, "task_passed": true, "parsed": { "cancelled_id": "R017", "cancelled_owner": "林青", "cancel_reason": "容量评估未通过", "approved_id": "R089", "rollback_threshold": "错误率超过2%" }, "recording_archive": "strata-3090-evidence.zip", "recording": "recordings/confirm-candidate-off/r1--long_read.json", "decode_expert_cache_hit_pct": 88.2, "expert_blob_file_reads": 0 }, { "group": "confirm-candidate-off", "id": "r2--code_retry_fix", "case_id": "code_retry", "complete": false, "finish_reason": "length", "ttft_s": 0.6348157239990542, "e2e_s": 77.5349321350077, "input_tokens": 343, "output_tokens": 1024, "reused_tokens": 280, "prefill_tps": 102.0, "decode_tps": 13.3, "decode_s": 76.91, "draft_accepted": 664, "draft_offered": 1084, "acceptance": 0.6125461254612546, "tail_skipped_experts": 0, "read_mb": 533.024768, "major_faults": 4382, "cpu_core_equiv": 1.3832474866070812, "output_sha256": "7ae3f1740e2d97f3d4f0b2548c79343ee5a8c9edef4b9af19d2e9cb5ca3c4024", "gpu_util_mean": 98.78947368421052, "gpu_power_mean_w": 181.65684210526317, "gpu_temperature_max": 68.0, "gpu_vram_peak_mib": 24095.0, "ram_available_min_gib": 11.231781005859375, "tests_passed": null, "tests_total": null, "tests_all_passed": false, "task_passed": false, "recording_archive": "strata-3090-evidence.zip", "recording": "recordings/confirm-candidate-off/r2--code_retry_fix.json", "decode_expert_cache_hit_pct": 97.5, "expert_blob_file_reads": 0 }, { "group": "confirm-candidate-off", "id": "r2--code_log", "case_id": "code_log", "complete": true, "finish_reason": "stop", "ttft_s": 0.5215514339943184, "e2e_s": 61.2905201829999, "input_tokens": 166, "output_tokens": 751, "reused_tokens": 0, "prefill_tps": 329.6, "decode_tps": 12.4, "decode_s": 60.7805, "draft_accepted": 526, "draft_offered": 676, "acceptance": 0.7781065088757396, "tail_skipped_experts": 0, "read_mb": 366.20288, "major_faults": 3009, "cpu_core_equiv": 1.2427696774733397, "output_sha256": "ee97ecee2167ad8188027c9a99b03ed5e24f2fb9953144621ac6a3e1412a2227", "gpu_util_mean": 97.6774193548387, "gpu_power_mean_w": 173.71774193548387, "gpu_temperature_max": 66.0, "gpu_vram_peak_mib": 24095.0, "ram_available_min_gib": 11.345470428466797, "tests_passed": 10, "tests_total": 10, "tests_all_passed": true, "task_passed": true, "recording_archive": "strata-3090-evidence.zip", "recording": "recordings/confirm-candidate-off/r2--code_log.json", "decode_expert_cache_hit_pct": 95.4, "expert_blob_file_reads": 0 }, { "group": "off-a", "id": "code_merge", "case_id": "code_merge", "complete": true, "finish_reason": "stop", "ttft_s": 29.440613753002253, "e2e_s": 66.29333852800482, "input_tokens": 117, "output_tokens": 160, "reused_tokens": 0, "prefill_tps": 4.0, "decode_tps": 4.3, "decode_s": 36.8658, "draft_accepted": 110, "draft_offered": 128, "acceptance": 0.859375, "tail_skipped_experts": 0, "read_mb": 344.7808, "major_faults": 2800, "cpu_core_equiv": 0.725864786243527, "output_sha256": "94e9d577b8fca71083a22ae5aa7bb15caff8843c29c1f20e7b074a073b7963f1", "gpu_util_mean": 55.125, "gpu_power_mean_w": 125.1128125, "gpu_temperature_max": 62.0, "gpu_vram_peak_mib": 24085.0, "ram_available_min_gib": 12.309764862060547, "tests_passed": 67, "tests_total": 67, "tests_all_passed": true, "task_passed": true, "recording_archive": "strata-3090-evidence.zip", "recording": "recordings/off-a/code_merge.json", "decode_expert_cache_hit_pct": 86.7, "expert_blob_file_reads": 0 }, { "group": "off-b", "id": "code_merge", "case_id": "code_merge", "complete": true, "finish_reason": "stop", "ttft_s": 0.42451930900278967, "e2e_s": 2.748027740002726, "input_tokens": 117, "output_tokens": 160, "reused_tokens": 0, "prefill_tps": 286.1, "decode_tps": 68.6, "decode_s": 2.3314, "draft_accepted": 114, "draft_offered": 138, "acceptance": 0.8260869565217391, "tail_skipped_experts": 0, "read_mb": 1.961984, "major_faults": 16, "cpu_core_equiv": 12.765518880812024, "output_sha256": "94e9d577b8fca71083a22ae5aa7bb15caff8843c29c1f20e7b074a073b7963f1", "gpu_util_mean": 100.0, "gpu_power_mean_w": 256.73, "gpu_temperature_max": 64.0, "gpu_vram_peak_mib": 24087.0, "ram_available_min_gib": 12.150337219238281, "tests_passed": 67, "tests_total": 67, "tests_all_passed": true, "task_passed": true, "recording_archive": "strata-3090-evidence.zip", "recording": "recordings/off-b/code_merge.json", "decode_expert_cache_hit_pct": 82.5, "expert_blob_file_reads": 0 }, { "group": "sweep-p2048-w7-s4", "id": "sweep-0-mtp07-code_merge", "case_id": "code_merge", "complete": true, "finish_reason": "stop", "ttft_s": 0.4360485250072088, "e2e_s": 1.4955128069996135, "input_tokens": 117, "output_tokens": 160, "reused_tokens": 0, "prefill_tps": 277.6, "decode_tps": 149.7, "decode_s": 1.0691, "draft_accepted": 113, "draft_offered": 128, "acceptance": 0.8828125, "tail_skipped_experts": 0, "read_mb": 0.0, "major_faults": 0, "cpu_core_equiv": 6.927389689684333, "output_sha256": "94e9d577b8fca71083a22ae5aa7bb15caff8843c29c1f20e7b074a073b7963f1", "gpu_util_mean": 95.0, "gpu_power_mean_w": 253.32, "gpu_temperature_max": 69.0, "gpu_vram_peak_mib": 24085.0, "ram_available_min_gib": 11.390975952148438, "tests_passed": 67, "tests_total": 67, "tests_all_passed": true, "task_passed": true, "recording_archive": "strata-3090-evidence.zip", "recording": "recordings/sweep-p2048-w7-s4/sweep-0-mtp07-code_merge.json", "decode_expert_cache_hit_pct": 97.1, "expert_blob_file_reads": 0 } ], "prefill_ablation": [ { "chunk": 256, "share": "auto", "n": 6, "fields_all_correct": 6, "zero_reuse": true, "ttft_s": { "mean": 13.281485168006233, "median": 13.221309856002335, "min": 11.671712854004, "max": 14.909158762020525 }, "prefill_ms": { "mean": 13248.5, "median": 13182.0, "min": 11650.0, "max": 14868.0 }, "prefill_tps": { "mean": 335.01666666666665, "median": 334.85, "min": 295.4, "max": 377.0 }, "io_MB": { "mean": 0.0, "median": 0.0, "min": 0.0, "max": 0.0 }, "cpu_seconds": { "mean": 63.321666666666665, "median": 63.475, "min": 45.5, "max": 82.58 }, "e2e_s": { "mean": 14.119787230001142, "median": 13.987620906002121, "min": 12.49839396198513, "max": 15.799006858986104 }, "blocks": [ { "group": "01-p256-auto", "ttft_mean": 11.953733112013046, "io_MB_max": 0.0 }, { "group": "08-p256-auto", "ttft_mean": 14.60923722399942, "io_MB_max": 0.0 } ] }, { "chunk": 256, "share": "off", "n": 6, "fields_all_correct": 6, "zero_reuse": true, "ttft_s": { "mean": 16.113288147826097, "median": 14.021417344993097, "min": 11.716514083993388, "max": 26.140193678991636 }, "prefill_ms": { "mean": 16085.833333333334, "median": 13993.0, "min": 11675.0, "max": 26119.0 }, "prefill_tps": { "mean": 291.71666666666664, "median": 313.9, "min": 168.2, "max": 376.2 }, "io_MB": { "mean": 0.0, "median": 0.0, "min": 0.0, "max": 0.0 }, "cpu_seconds": { "mean": 22.955000000000002, "median": 21.205, "min": 18.45, "max": 33.84 }, "e2e_s": { "mean": 17.220172933656915, "median": 15.21962438848277, "min": 12.844750733987894, "max": 27.286243626003852 }, "blocks": [ { "group": "02-p256-off", "ttft_mean": 18.29183800199341, "io_MB_max": 0.0 }, { "group": "07-p256-off", "ttft_mean": 13.934738293658787, "io_MB_max": 0.0 } ] }, { "chunk": 1024, "share": "off", "n": 6, "fields_all_correct": 6, "zero_reuse": true, "ttft_s": { "mean": 6.1358481408339385, "median": 5.65961030300241, "min": 5.069840067997575, "max": 8.557597228005761 }, "prefill_ms": { "mean": 6108.166666666667, "median": 5629.5, "min": 5042.0, "max": 8536.0 }, "prefill_tps": { "mean": 743.9666666666667, "median": 781.8499999999999, "min": 514.5, "max": 871.1 }, "io_MB": { "mean": 0.0, "median": 0.0, "min": 0.0, "max": 0.0 }, "cpu_seconds": { "mean": 17.538333333333334, "median": 17.29, "min": 13.32, "max": 21.62 }, "e2e_s": { "mean": 7.092880115170071, "median": 6.7507419975008816, "min": 5.645317306014476, "max": 9.16050857200753 }, "blocks": [ { "group": "03-p1024-off", "ttft_mean": 5.203280210999462, "io_MB_max": 0.0 }, { "group": "06-p1024-off", "ttft_mean": 7.068416070668415, "io_MB_max": 0.0 } ] }, { "chunk": 2048, "share": "off", "n": 6, "fields_all_correct": 6, "zero_reuse": true, "ttft_s": { "mean": 3.5422005238360725, "median": 3.547208837015205, "min": 3.400591968995286, "max": 3.6359843389946036 }, "prefill_ms": { "mean": 3504.6666666666665, "median": 3512.5, "min": 3364.0, "max": 3601.0 }, "prefill_tps": { "mean": 1253.8166666666666, "median": 1250.4, "min": 1219.7, "max": 1305.7 }, "io_MB": { "mean": 0.0, "median": 0.0, "min": 0.0, "max": 0.0 }, "cpu_seconds": { "mean": 11.698333333333334, "median": 11.745000000000001, "min": 10.38, "max": 13.24 }, "e2e_s": { "mean": 4.399800813999415, "median": 4.361849003005773, "min": 4.113352790998761, "max": 4.644212639017496 }, "blocks": [ { "group": "04-p2048-off", "ttft_mean": 3.4895625696711554, "io_MB_max": 0.0 }, { "group": "05-p2048-off", "ttft_mean": 3.5948384780009897, "io_MB_max": 0.0 } ] } ], "startup_pcie_probes": [ { "group": "01-p256-auto", "recording_archive": "prefill-ablation-evidence.zip", "startup_h2d_GB_s": 16.1, "initial_pcie_frac": 0.44, "recording": "recordings/01-p256-auto/engine.log" }, { "group": "02-p256-off", "recording_archive": "prefill-ablation-evidence.zip", "startup_h2d_GB_s": 16.5, "initial_pcie_frac": 0.45, "recording": "recordings/02-p256-off/engine.log" }, { "group": "03-p1024-off", "recording_archive": "prefill-ablation-evidence.zip", "startup_h2d_GB_s": 3.8, "initial_pcie_frac": 0.1, "recording": "recordings/03-p1024-off/engine.log" }, { "group": "04-p2048-off", "recording_archive": "prefill-ablation-evidence.zip", "startup_h2d_GB_s": 19.7, "initial_pcie_frac": 0.54, "recording": "recordings/04-p2048-off/engine.log" }, { "group": "05-p2048-off", "recording_archive": "prefill-ablation-evidence.zip", "startup_h2d_GB_s": 16.8, "initial_pcie_frac": 0.46, "recording": "recordings/05-p2048-off/engine.log" }, { "group": "06-p1024-off", "recording_archive": "prefill-ablation-evidence.zip", "startup_h2d_GB_s": 14.5, "initial_pcie_frac": 0.4, "recording": "recordings/06-p1024-off/engine.log" }, { "group": "07-p256-off", "recording_archive": "prefill-ablation-evidence.zip", "startup_h2d_GB_s": 16.9, "initial_pcie_frac": 0.46, "recording": "recordings/07-p256-off/engine.log" }, { "group": "08-p256-auto", "recording_archive": "prefill-ablation-evidence.zip", "startup_h2d_GB_s": 22.8, "initial_pcie_frac": 0.55, "recording": "recordings/08-p256-auto/engine.log" } ], "methods": { "read_mb": "Delta /proc/PID/io read_bytes / 1e6 over request; includes prefill and decode, excludes model load", "major_faults": "Delta /proc/PID/stat field 12 (majflt), process only, excluding child faults", "cpu_seconds": "Delta process utime+stime divided by CLK_TCK=100 on tested Linux; all threads, both stages", "gpu": "nvidia-smi util/memory/power/temp/clocks sampled every 2 seconds and filtered by request interval", "ram_available": "Minimum /proc/meminfo MemAvailable among samples inside request interval", "acceptance": "Sum draft_n_accepted / sum draft_n within the stated task set", "startup_h2d": "Engine startup probe, best of four; initialization/calibration only", "pcie_activity": "Reverse groups only, nvidia-smi dmon -s t -d 1 -o DT; local timestamps UTC+08:00; sampled rates, not total bytes" } }