Skip to content

fix(grpo): apply effort shaping to async Gym rollouts - #3897

Merged
yfw merged 1 commit into
super-v3.5-posttrainingfrom
yifu/cherry-pick-pr3885-super-v3.5
Aug 28, 2026
Merged

fix(grpo): apply effort shaping to async Gym rollouts#3897
yfw merged 1 commit into
super-v3.5-posttrainingfrom
yifu/cherry-pick-pr3885-super-v3.5

Conversation

@yfw

@yfw yfw commented Aug 28, 2026

Copy link
Copy Markdown
Contributor

Summary

Source commit: fbc6a8811835747e9e602df1df26288186a26eb8

Validation

  • Async retry wiring and effort-shaping unit coverage — 10 passed
  • pre-commit run --all-files — passed
  • git diff --check — passed

Legacy async GRPO forwarded reward penalties to NeMo Gym rollout postprocessing but omitted the configured effort levels. As a result, training silently skipped length-based effort shaping even though synchronous rollouts and validation applied it.

Resolve the existing validated effort config in the async trajectory collector and pass it to every NeMo Gym stream attempt. Add regression coverage at the collector-to-rollout boundary, including the retry path.

Signed-off-by: Yi-Fu Wu <yifu.wu@gmail.com>
(cherry picked from commit fbc6a88)
@yfw
yfw requested review from a team as code owners August 28, 2026 16:31
@copy-pr-bot

copy-pr-bot Bot commented Aug 28, 2026

Copy link
Copy Markdown

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

@yfw
yfw merged commit 20ae0bb into super-v3.5-posttraining Aug 28, 2026
10 checks passed
@yfw
yfw deleted the yifu/cherry-pick-pr3885-super-v3.5 branch August 28, 2026 16:36
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