bench-e1b-w8-drop.txt 1009 B

123456789
  1. c1-prose decode 14.53 tok/s (min 14.50 max 14.54 sd 0.01) ttft 0.298s toks [512, 512, 512, 512, 512] distinct 4 err 0 2026-09-24T01:27:50.644892Z..2026-09-24T01:30:48.033711Z
  2. c1-code decode 14.29 tok/s (min 14.25 max 14.32 sd 0.02) ttft 0.301s toks [512, 512, 512, 512, 512] distinct 2 err 0 2026-09-24T01:30:48.034053Z..2026-09-24T01:33:48.373670Z
  3. c1-chat_sampled decode 14.29 tok/s (min 14.25 max 14.40 sd 0.06) ttft 0.277s toks [512, 512, 512, 512, 512] distinct 5 err 0 2026-09-24T01:33:48.373919Z..2026-09-24T01:36:48.226769Z
  4. c4 agg 26.60 tok/s (min 26.17 max 26.74) per-stream 6.70 2026-09-24T01:36:48.227020Z..2026-09-24T01:43:13.987521Z
  5. c1-prose accept_len 1.02 step 71.3 ms (n=63 intervals, bs=1)
  6. c1-code accept_len 1.00 step 70.6 ms (n=64 intervals, bs=1.0)
  7. c1-chat_sampled accept_len 1.00 step 71.1 ms (n=63 intervals, bs=1)
  8. c4 accept_len 1.02 step 153.2 ms (n=64 intervals, bs=4.0)
  9. head full-stall us during bench: 15829