Skip to content

fix: support DDP online rollout - #29

Closed
0z5a wants to merge 1 commit into
ViperEkura:mainfrom
0z5a:codex/fix-ddp-online-rollout
Closed

fix: support DDP online rollout#29
0z5a wants to merge 1 commit into
ViperEkura:mainfrom
0z5a:codex/fix-ddp-online-rollout

Conversation

@0z5a

@0z5a 0z5a commented Sep 2, 2026

Copy link
Copy Markdown
Contributor

Summary

  • expose executor-owned training and inference model views
  • unwrap DDP for rollout inference and keep CPU DDP device handling valid
  • fail fast for unsupported FSDP and compiled online rollout combinations

Validation

  • CPU DDP focused tests: 11 passed
  • GPU5 focused DDP rollout tests: 2 passed
  • bash scripts/pre_commit.sh --skip-deps
  • 598 passed, 50 skipped

- Expose executor-owned training and inference model views and unwrap DDP for the scheduler
- Fail fast for unsupported FSDP and torch.compile rollout configurations
- Cover single-process and two-rank online GRPO and DPO execution
@0z5a

0z5a commented Sep 2, 2026

Copy link
Copy Markdown
Contributor Author

Superseded by #38. The original commit author was incorrectly attributed to ViperEkura by a repository-local Git config override. The replacement has the identical tree and is correctly attributed to 0z5a.

@0z5a 0z5a closed this Sep 2, 2026
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.

2 participants