Skip to content

Conversation

@cklxx
Copy link
Owner

@cklxx cklxx commented Dec 22, 2025

Summary

  • refactor OPSM mask computation to be context-parallel aware without duplicating logic in the loss
  • restore gradient connectivity when policy log-prob tensors are empty to avoid distributed hangs
  • validate rollout buffer group sizes to prevent silently misaligned sample counts

Testing

  • python -m ruff check slime slime_plugins tests
  • python -m pytest (fails: missing optional dependencies such as flash_attn and model-specific test modules)

Codex Task

@cklxx cklxx merged commit d398b45 into codex/optimize-training-time-for-context-parallelism Dec 22, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant