Skip to content

Fix FSDP bug when parameters are partially frozen; Support saving ckpt with custom dtype - #226

Open
bfshi wants to merge 5 commits into
ByteDance-Seed:mainfrom
bfshi:main
Open

Fix FSDP bug when parameters are partially frozen; Support saving ckpt with custom dtype#226
bfshi wants to merge 5 commits into
ByteDance-Seed:mainfrom
bfshi:main

Conversation

@bfshi

@bfshi bfshi commented Aug 4, 2025

Copy link
Copy Markdown
  1. Fix FSDP bug of not able to train with partially frozen parameters
  2. Fix bug of not stopping training after total_steps training steps and not saving the ckpt when training ends
  3. Add a save_dtype arg to support saving ckpt with custom dtype.
  4. Minor fix of missing detail in TRAIN.md

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant