Historical tool output | archived output | JSONL line 7824 | 2026-09-08T11:28:32.912Z ##### 20260907-055513-v1-baseline.txt T=0.0: TTFT=0.16s decode(median of 3)=28.5 tok/s tokens=[300, 300, 300] T=0.7: TTFT=0.25s decode(median of 3)=19.4 tok/s tokens=[300, 300, 285] --- 2) concurrency sweep (200 tokens each, T=0.7) --- N= 1 wall= 12.0s tokens= 200 aggregate= 16.7 tok/s per-stream= 16.7 tok/s N= 2 wall= 16.3s tokens= 400 aggregate= 24.6 tok/s per-stream= 12.3 tok/s N= 4 wall= 23.3s tokens= 800 aggregate= 34.4 tok/s per-stream= 8.6 tok/s N= 8 wall= 43.5s tokens= 1600 aggregate= 36.7 tok/s per-stream= 4.6 tok/s --- 3) MTP acceptance --- ##### 20260907-111529-v2-262k-cmode0.txt T=0.0: TTFT=0.23s decode(median of 3)=24.2 tok/s tokens=[300, 300, 300] T=0.7: TTFT=0.14s decode(median of 3)=23.8 tok/s tokens=[300, 300, 300] --- 2) concurrency sweep (200 tokens each, T=0.7) --- N= 1 wall= 8.1s tokens= 200 aggregate= 24.6 tok/s per-stream= 24.6 tok/s N= 2 wall= 12.5s tokens= 400 aggregate= 31.9 tok/s per-stream= 16.0 tok/s N= 4 wall= 14.1s tokens= 800 aggregate= 56.8 tok/s per-stream= 14.2 tok/s N= 8 wall= 17.3s tokens= 1600 aggregate= 92.7 tok/s per-stream= 11.6 tok/s --- 3) MTP acceptance --- drafts=2663 draft_tokens=7989 accepted=2732 acceptance=34.2% mean_accepted_per_step=1.03 per-position=['58%', '31%', '13%'] ##### 20260907-115835-mp-262k.txt T=0.0: TTFT=0.23s decode(median of 3)=20.6 tok/s tokens=[286, 286, 286] T=0.7: TTFT=0.24s decode(median of 3)=20.8 tok/s tokens=[273, 300, 289] --- 2) concurrency sweep (200 tokens each, T=0.7) --- N= 1 wall= 10.6s tokens= 200 aggregate= 18.8 tok/s per-stream= 18.8 tok/s N= 2 wall= 12.7s tokens= 400 aggregate= 31.6 tok/s per-stream= 15.8 tok/s N= 4 wall= 16.4s tokens= 800 aggregate= 48.7 tok/s per-stream= 12.2 tok/s N= 8 wall= 17.8s tokens= 1600 aggregate= 89.9 tok/s per-stream= 11.2 tok/s --- 3) MTP acceptance --- drafts=2704 draft_tokens=8112 accepted=2607 acceptance=32.1% mean_accepted_per_step=0.96 per-position=['56%', '27%', '13%'] ##### 20260907-123927-v2-262k-restored.txt T=0.0: TTFT=0.22s decode(median of 3)=23.3 tok/s tokens=[286, 286, 286] T=0.7: TTFT=0.22s decode(median of 3)=23.1 tok/s tokens=[273, 300, 289] --- 2) concurrency sweep (200 tokens each, T=0.7) --- N= 1 wall= 9.1s tokens= 200 aggregate= 22.0 tok/s per-stream= 22.0 tok/s N= 2 wall= 11.8s tokens= 400 aggregate= 33.8 tok/s per-stream= 16.9 tok/s N= 4 wall= 15.3s tokens= 800 aggregate= 52.2 tok/s per-stream= 13.1 tok/s N= 8 wall= 17.1s tokens= 1600 aggregate= 93.5 tok/s per-stream= 11.7 tok/s --- 3) MTP acceptance --- drafts=2705 draft_tokens=8115 accepted=2595 acceptance=32.0% mean_accepted_per_step=0.96 per-position=['56%', '27%', '13%'] ##### 20260907-212629-B2-hybrid.txt T=0.0: TTFT=0.13s decode(median of 3)=27.3 tok/s tokens=[300, 300, 300] T=0.7: TTFT=0.13s decode(median of 3)=26.3 tok/s tokens=[300, 300, 292] --- 2) concurrency sweep (200 tokens each, T=0.7) --- N= 1 wall= 7.6s tokens= 200 aggregate= 26.2 tok/s per-stream= 26.2 tok/s N= 2 wall= 10.9s tokens= 400 aggregate= 36.6 tok/s per-stream= 18.3 tok/s N= 4 wall= 13.4s tokens= 800 aggregate= 59.7 tok/s per-stream= 14.9 tok/s N= 8 wall= 17.3s tokens= 1600 aggregate= 92.4 tok/s per-stream= 11.5 tok/s --- 3) MTP acceptance --- drafts=2650 draft_tokens=7950 accepted=2730 acceptance=34.3% mean_accepted_per_step=1.03 per-position=['58%', '30%', '15%'] ##### 20260908-010234-C2-mtp0.txt T=0.0: TTFT=0.12s decode(median of 3)=23.7 tok/s tokens=[300, 300, 300] T=0.7: TTFT=0.12s decode(median of 3)=22.4 tok/s tokens=[300, 300, 300] --- 2) concurrency sweep (200 tokens each, T=0.7) --- N= 1 wall= 10.2s tokens= 200 aggregate= 19.7 tok/s per-stream= 19.7 tok/s N= 2 wall= 12.8s tokens= 400 aggregate= 31.1 tok/s per-stream= 15.6 tok/s N= 4 wall= 12.8s tokens= 800 aggregate= 62.7 tok/s per-stream= 15.7 tok/s N= 8 wall= 16.3s tokens= 1600 aggregate= 98.3 tok/s per-stream= 12.3 tok/s --- 3) MTP acceptance --- drafts=1 draft_tokens=0 accepted=0 acceptance=0.0% mean_accepted_per_step=0.00 per-position=[] ##### 20260908-051613-E-nvidia-ckpt.txt T=0.0: TTFT=0.13s decode(median of 3)=24.2 tok/s tokens=[300, 300, 300] T=0.7: TTFT=0.14s decode(median of 3)=23.8 tok/s tokens=[300, 299, 300] --- 2) concurrency sweep (200 tokens each, T=0.7) --- N= 1 wall= 8.6s tokens= 200 aggregate= 23.1 tok/s per-stream= 23.1 tok/s N= 2 wall= 11.5s tokens= 400 aggregate= 34.8 tok/s per-stream= 17.4 tok/s N= 4 wall= 13.4s tokens= 800 aggregate= 59.8 tok/s per-stream= 15.0 tok/s N= 8 wall= 16.7s tokens= 1600 aggregate= 95.8 tok/s per-stream= 12.0 tok/s --- 3) MTP acceptance --- drafts=2692 draft_tokens=8076 accepted=2691 acceptance=33.3% mean_accepted_per_step=1.00 per-position=['57%', '29%', '14%'] ##### 20260908-061035-F-nvidia-hybrid.txt 中文散文 解码 23.7 tok/s TTFT 0.20s (300 token) 接受率 nan% 每步接受 nan 按位置 ['33500%', '16100%', '8000%'] Python代码 解码 37.5 tok/s TTFT 0.15s (300 token) 接受率 nan% 每步接受 nan 按位置 ['33600%', '28000%', '20800%'] 重复结构 解码 49.3 tok/s TTFT 0.23s (300 token) 接受率 nan% 每步接受 nan 按位置 ['30400%', '30400%', '29200%'] JSON输出 解码 43.2 tok/s TTFT 0.27s (300 token) 接受率 nan% 每步接受 nan 按位置 ['31200%', '30400%', '26400%'] ##### 20260908-062513-F-nvidia-hybrid.txt 中文散文 解码 22.3 tok/s TTFT 0.14s (300 token) 接受率 29.0% 每步接受 0.87 按位置 ['51%', '24%', '11%'] Python代码 解码 38.9 tok/s TTFT 0.15s (300 token) 接受率 71.5% 每步接受 2.15 按位置 ['88%', '73%', '54%'] 重复结构 解码 48.9 tok/s TTFT 0.23s (300 token) 接受率 98.7% 每步接受 2.96 按位置 ['100%', '100%', '96%'] JSON输出 解码 46.2 tok/s TTFT 0.24s (300 token) 接受率 92.8% 每步接受 2.78 按位置 ['99%', '96%', '84%'] ##### 20260908-065400-G-radixark-hybrid.txt 中文散文 解码 25.7 tok/s TTFT 0.28s (300 token) 接受率 36.4% 每步接受 1.09 按位置 ['62%', '31%', '17%'] Python代码 解码 38.5 tok/s TTFT 0.23s (300 token) 接受率 73.8% 每步接受 2.22 按位置 ['89%', '74%', '58%'] 重复结构 解码 49.8 tok/s TTFT 0.23s (300 token) 接受率 97.8% 每步接受 2.93 按位置 ['100%', '99%', '95%'] JSON输出 解码 47.9 tok/s TTFT 0.26s (300 token) 接受率 94.4% 每步接受 2.83 按位置 ['98%', '95%', '90%'] ##### 20260908-071115-H-radixark-hybrid-b2048.txt 中文散文 解码 26.9 tok/s TTFT 0.22s (300 token) 接受率 36.4% 每步接受 1.09 按位置 ['62%', '31%', '17%'] Python代码 解码 40.5 tok/s TTFT 0.23s (300 token) 接受率 74.4% 每步接受 2.23 按位置 ['89%', '75%', '59%'] 重复结构 解码 48.8 tok/s TTFT 0.16s (300 token) 接受率 98.1% 每步接受 2.94 按位置 ['100%', '99%', '96%'] JSON输出 解码 48.3 tok/s TTFT 0.17s (300 token) 接受率 94.1% 每步接受 2.82 按位置 ['97%', '95%', '90%'] ##### 20260908-071455-I-final-radixark-hybrid-b2048.txt T=0.0: TTFT=0.21s decode(median of 3)=26.8 tok/s tokens=[300, 300, 300] T=0.7: TTFT=0.21s decode(median of 3)=24.1 tok/s tokens=[291, 278, 300] --- 2) concurrency sweep (200 tokens each, T=0.7) --- N= 1 wall= 7.9s tokens= 200 aggregate= 25.4 tok/s per-stream= 25.4 tok/s N= 2 wall= 11.4s tokens= 400 aggregate= 35.2 tok/s per-stream= 17.6 tok/s N= 4 wall= 13.2s tokens= 800 aggregate= 60.8 tok/s per-stream= 15.2 tok/s N= 8 wall= 17.5s tokens= 1600 aggregate= 91.6 tok/s per-stream= 11.4 tok/s --- 3) MTP acceptance --- drafts=4639 draft_tokens=13917 accepted=6709 acceptance=48.2% mean_accepted_per_step=1.45 per-position=['68%', '45%', '32%']