Skip to content

Worker: 0.1.745

Choose a tag to compare

View all tags
@luke-lombardi luke-lombardi released this 07 Sep 02:47
worker-0.1.745
5f3e354

Worker fix on top of 0.1.744:

  • clip: parallel-pull every layer >= 32 MiB in 32 MiB ranges and resume ranges that fall far behind their peers on a new connection (#1876, beam-cloud/clip@9420edd). Remote (skynet/OVH) workers pulling from ECR were bounded by whichever S3 front-end a single flow landed on: a 3 GiB model layer took 180 s and a 400 MiB torch layer 249 s at 1-3 MiB/s while the link could do 200+ MiB/s. Measured on a skynet worker: 2.9 GiB layer 15 s (was 37-473 s), 383 MiB layer 4-6 s (was 5-108 s).
  • clip: front-ends that starved a range are avoided for 2 min; per-layer info log parallel registry blob prefetch complete with duration, MiB/s, straggler_aborts, retries; disk-headroom skip is now a warning.

Gateway unchanged (0.1.768).

Assets 2
Loading