8000 v0.16.0: native verl Hydra input and deterministic sampling by DaoyuanLi2816 · Pull Request #123 · DaoyuanLi2816/mini-verl · GitHub
[go: up one dir, main page]
Skip to content

v0.16.0: native verl Hydra input and deterministic sampling - #123

Merged
DaoyuanLi2816 merged 4 commits into
mainfrom
v0.16-native-hydra-semantics
Sep 11, 2026
Merged

v0.16.0: native verl Hydra input and deterministic sampling#123
DaoyuanLi2816 merged 4 commits into
mainfrom
v0.16-native-hydra-semantics

Conversation

@DaoyuanLi2816
@DaoyuanLi2816 DaoyuanLi2816 commented Sep 11, 2026 •
edited
Loading
Copy link
Copy Markdown
Owner

Summary

  • Compose pinned verl v0.9.0 config trees, names, defaults and ordinary Hydra overrides directly, with immutable input and resolved-byte provenance.
  • Add the v5 profile for upstream-conformant actor/critic shuffled epochs, reward-variance group filtering and deterministic replacement. Preserve v1-v4 identities.
  • Separate composition, semantic acceptance and hardware feasibility; retain explicit runtime bindings and precise rejection paths.
  • Add exact-candidate native PPO/GRPO GPU qualification, resume and handoff gates to the existing one-build OIDC release chain.

Evidence and boundaries

  • Native corpus: 16 cases; 14 compose, 12 are semantically accepted, 2 are semantic rejections and 2 reject composition. Static acceptance does not imply production-model execution.
  • Installed-wheel development rehearsal: RTX 4080, PPO 4 actor/4 critic updates, GRPO 4 actor updates with 4 filtered groups, exact tensor replay and inspected handoff. Final release requires fresh full qualification of the merged candidate bytes.
  • Refill follows pinned V1 behavior; the unused legacy over_sample_rate is informational. Equal-age selection uses a documented local dispatch-order tie-break. No distributed execution or model-quality claim.
  • Frozen results and historical v4 implementation remain byte-identical.

Validation

  • Final local CPU suite: 2814 passed, 10 platform skips, 24 deselected; combined line/branch coverage 82.15%.
  • GPU tests: 10 passed; network tests: 16 passed; exact upstream conformance included.
  • Ruff, mypy, actionlint, strict docs, links, generated artifacts, package build and twine passed.
  • Four CI browser viewports manually reviewed; screenshots attached to docs run 34577764046.
  • Green PR checks: CI 34577764043, build and extracted-sdist/clean installs 34577764158, docs 34577764046, pinned bridge 34577764059.
  • Squash merge 7a6ad7d has the identical tested source tree.

Release gate

Publish only after fresh exact merged-SHA full GPU qualification and the non-publishing release rehearsal. Development rehearsal is not release evidence.

@DaoyuanLi2816
DaoyuanLi2816 merged commit 7a6ad7d into main Sep 11, 2026
16 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

4CA8 No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@DaoyuanLi2816
0