8000 Productize local PPO and GRPO workflows for v0.14.0 by DaoyuanLi2816 · Pull Request #118 · DaoyuanLi2816/mini-verl · GitHub
[go: up one dir, main page]

Skip to content

Productize local PPO and GRPO workflows for v0.14.0 - #118

Merged
DaoyuanLi2816 merged 5 commits into
mainfrom
v0.14.0-p0-productization
Sep 9, 2026
Merged

DaoyuanLi2816 merged 5 commits into
mainfrom
v0.14.0-p0-productization

Conversation

@DaoyuanLi2816
@DaoyuanLi2816 DaoyuanLi2816 commented Sep 9, 2026 •
Copy link
Copy Markdown
Owner

Summary

  • Package installed PPO/GRPO import, inspect, resume and handoff workflows with upstream adaptation provenance.
  • Add six upstream launch sources / seven Hydra-resolved cases and immutable v3 compiler with correct prompt-minibatch units.
  • Bind PPO role identities and restore checkpoint-consistent logs after interrupted updates.
  • Add strict RL inspection and exact-wheel product qualification gates; foreground local RL in English/Chinese documentation.

Validation

  • 2,655 non-GPU/non-network tests passed, 83% branch coverage; 14 documented platform/source skips.
  • 10 local RTX 4080 GPU tests, 16 network tests and four exact-source upstream numerical conformance tests passed.
  • Ruff, format, mypy, actionlint, generated corpus/examples, strict docs and four-viewport browser checks passed (44 rendered SVG instances).
  • Development wheel installation exercised both real PPO/GRPO workflows with exact tensor replay. Formal release evidence must qualify the final release SHA and accepted candidate bytes.
  • Final metadata is 0.14.0; merge only after all checks pass, then run exact-SHA full qualification and the non-publishing release gate before tagging.

All frozen scientific results remain byte-identical. The seven complete upstream configurations are rejected with field-level reasons, not represented as executed experiments. The two packaged workflows are explicit local adaptations. Distributed execution is not part of this release.

@DaoyuanLi2816
DaoyuanLi2816 merged commit 60afc63 into main Sep 9, 2026
16 checks passed
@DaoyuanLi2816
DaoyuanLi2816 deleted the v0.14.0-p0-productization branch September 9, 2026 06:12
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

0