8000 v0.15.0: run resolved verl configs directly on one GPU by DaoyuanLi2816 · Pull Request #120 · DaoyuanLi2816/mini-verl · GitHub
[go: up one dir, main page]

Skip to content

v0.15.0: run resolved verl configs directly on one GPU - #120

Merged
DaoyuanLi2816 merged 3 commits into
mainfrom
v0.15-direct-verl-config
Sep 10, 2026
Merged

DaoyuanLi2816 merged 3 commits into
mainfrom
v0.15-direct-verl-config

Conversation

@DaoyuanLi2816
@DaoyuanLi2816 DaoyuanLi2816 commented Sep 10, 2026 •
Copy link
Copy Markdown
Owner

Summary

  • Add the immutable v4 direct-config compiler and typed runtime bindings, with separate semantic, missing-input and hardware states.
  • Preserve upstream logical reductions, filtering, epoch schedules, repeated actor passes, independent critic/reference roles and classifier reward semantics.
  • Add the generated 12-case audit (all seven complete official cases semantically accepted), installed direct PPO/GRPO qualification and English/Chinese documentation.

Validation

  • Final local regression: 2,731 passed, 28 skipped, 24 deselected; branch coverage exceeds the 80% gate.
  • 18 pinned upstream scalar/gradient/schedule checks, 10 RTX 4080 GPU tests and 15 network tests passed. One local upstream-v0.8 network test was skipped; the dedicated pinned bridge CI passed.
  • Installed development-wheel PPO/GRPO rehearsals passed exact tensor resume and handoff. Formal release qualification will use the hosted candidate wheel from the merged commit.
  • Ruff, mypy, strict MkDocs, generated artifacts, Markdown/privacy checks, package build and Twine passed. Four viewport screenshots were manually inspected; 44 rendered SVG instances passed the browser gate.
  • Torch-free collection regression was fixed and all four core Python versions now pass CI.

Release integrity

Frozen result/task JSON and JSONL, historical v1/v2/v3 compiler identities, tags and public model revisions are unchanged. The full-size source-model corpus was not executed. Publication requires green CI and first-attempt full exact-candidate GPU qualification, followed by a non-publishing rehearsal of the same bytes. Distributed verl execution is not claimed.

@DaoyuanLi2816
DaoyuanLi2816 marked this pull request as ready for review September 10, 2026 04:17
@DaoyuanLi2816
DaoyuanLi2816 merged commit 50c464a into main Sep 10, 2026
16 checks passed
@DaoyuanLi2816
DaoyuanLi2816 deleted the v0.15-direct-verl-config branch September 10, 2026 04:18
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