Skip to content

v0.21.3

Latest

Choose a tag to compare

@githubnemo githubnemo released this 13 Nov 10:11
· 161 commits to main since this release
599e264

Release 0.21.3 containing:

  • #3963: Only gather adapter weights when saving a DTensor TP model

This is mostly to fix a performance penalty with huggingface/transformers#49316 as this makes Trainer call save_pretrained on every rank, and with peft 0.21.2 that gathers the whole sharded model instead of just the adapter.

What's Changed

Full Changelog: v0.21.2...v0.21.3