Repository navigation
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
- Release prep/0.21.3 by @BenjaminBossan in #4164
Full Changelog: v0.21.2...v0.21.3