Skip to content

Add typed verl v0.8 OPD config compiler - #58

Merged
DaoyuanLi2816 merged 1 commit into
mainfrom
v0.8.0-verl-opd-config
Aug 12, 2026
Merged

Add typed verl v0.8 OPD config compiler#58
DaoyuanLi2816 merged 1 commit into
mainfrom
v0.8.0-verl-opd-config

Conversation

@DaoyuanLi2816

Copy link
Copy Markdown
Owner

Summary

  • add the typed verl-opd-v0.8-single-gpu-v1 config surface pinned to verl v0.8.0 at 7aed6b230776f963fa09509c10d9c3a767d1102c
  • compile resolved YAML plus repeatable dotted --set overrides into a deterministic SHA-256-bound compatibility report
  • classify every supplied leaf as exact, semantically conformant, locally reinterpreted, derived, informational only, or unsupported
  • fail closed on PG OPD, task-reward mixtures, KL penalties, multiple generations/teachers, and distributed dimensions
  • add an offline miniverl bridge compile-opd smoke that loads no weights and does not claim inference-engine or runtime equivalence

Scope boundary

This is PR A, the config/compiler layer only. config semantics executable: true means the supplied field set is inside the documented profile. The command explicitly reports that runtime execution is not provided by the config-only command. Parquet loading, prompt rollout, loss conformance, plan/run UX, and export arrive in later independently green PRs.

Validation

  • non-GPU/non-network: 2,135 passed, 6 skipped, 21 deselected; 86.21% branch coverage
  • GPU: 8 passed on the local RTX 4080
  • network: 15 passed
  • ruff check and format check
  • mypy over src/miniverl and scripts
  • actionlint 1.7.12
  • strict MkDocs, Markdown links, text integrity, release-state and PyPI README checks
  • wheel/sdist build and twine check
  • frozen benchmark/evidence/preregistration diff against v0.7.1: clean
  • calculator JSON SHA-256 remains 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc

@DaoyuanLi2816
DaoyuanLi2816 merged commit 87bea41 into main Aug 12, 2026
16 checks passed
@DaoyuanLi2816
DaoyuanLi2816 deleted the v0.8.0-verl-opd-config branch August 12, 2026 06:06
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