Skip to content

Closed by author#9376

Closed
yyswhsccc wants to merge 1 commit into
modelscope:mainfrom
yyswhsccc:bounty-radar/issue-9306-zero3-rollout-degeneration
Closed

Closed by author#9376
yyswhsccc wants to merge 1 commit into
modelscope:mainfrom
yyswhsccc:bounty-radar/issue-9306-zero3-rollout-degeneration

Conversation

@yyswhsccc
Copy link
Copy Markdown

@yyswhsccc yyswhsccc commented May 18, 2026

No description provided.

Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review

This pull request updates the _load_state_dict_to_vllm method in swift/rlhf_trainers/rollout_mixin.py to include a synchronization step. By importing and calling synchronize() after loading weights into the vLLM model, the change ensures that ZeRO-3 gathered tensors remain valid until all queued device copies are completed. I have no feedback to provide as there were no review comments.

@yyswhsccc yyswhsccc changed the title [bugfix] synchronize colocate vLLM weight loads Closed by author May 18, 2026
@yyswhsccc yyswhsccc closed this May 18, 2026
@yyswhsccc yyswhsccc deleted the bounty-radar/issue-9306-zero3-rollout-degeneration branch May 18, 2026 17:34
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