8000 Harden exact-SHA GPU qualification and maintainer release gates by DaoyuanLi2816 · Pull Request #85 · DaoyuanLi2816/mini-verl · GitHub
[go: up one dir, main page]

Skip to content

Harden exact-SHA GPU qualification and maintainer release gates - #85

Merged
DaoyuanLi2816 merged 9 commits into
mainfrom
v0.10.1-product-maturity
Aug 16, 2026
Merged

DaoyuanLi2816 merged 9 commits into
mainfrom
v0.10.1-product-maturity

Conversation

@DaoyuanLi2816
@DaoyuanLi2816 DaoyuanLi2816 commented Aug 15, 2026 •
Copy link
Copy Markdown
Owner

Summary

  • build one strict hosted-runner wheel/sdist candidate, then install and qualify those exact bytes on the private RTX 4080 runner
  • bind source SHA, workflow run and attempt, candidate manifest, wheel, environment, models and every declared evidence file into the qualification
  • reject cross-origin credential forwarding, evidence size drift, cross-attempt pairs and GitHub workflow reruns
  • require same-run candidate-distributions plus gpu-full-qualification for formal release; retain the complete full evidence set and checksums in workflow and GitHub Release assets
  • make PyPI publication idempotent only when the already-public wheel and sdist set exactly matches the candidate; partial, extra or mismatched releases fail closed

Root cause and contract

The prior path could forward API credentials across artifact redirects, validated artifact hashes without their declared byte counts, allowed rerun attempts to mix lifecycle boundaries, and accepted diagnostic smoke evidence for formal publication. The release chain is now:

fresh dispatch attempt 1 → hosted candidate bytes → same-run RTX 4080 full qualification → dry-run consumes identical artifacts → tag release publishes identical bytes and full evidence

No job rebuilds the accepted distributions. gpu-release-smoke remains useful diagnostics but cannot authorize publication.

Validation

  • exact PR head: 9809209479b44d4c413b71d4ad80fb192d86fa2e
  • focused release/security tests: 45 passed, including 301/302/303/307/308 header behavior and a two-origin local HTTP redirect
  • full non-GPU/non-network suite: 2362 passed, 8 skipped, 22 deselected; 84% branch coverage
  • extracted-sdist suite: 2359 passed, 11 skipped, 22 deselected
  • local RTX 4080 GPU markers: 8 passed; network markers: 16 passed
  • Ruff check/format, mypy, actionlint, strict MkDocs, Markdown links and git diff --check passed
  • clean core wheel install kept Torch absent; clean [train] install completed the real demo
  • development wheel SHA-256: 8e4b8d6eca11b04326e1d107ce4120a3049ea4c065aab77838c6efb55efa5201
  • development sdist SHA-256: 0e7dc2936b2677e5e4fd856c0682a11419b779cf553be581c585cfb5bcea855e
  • frozen calculator benchmark remains 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc; no frozen result JSON/JSONL changed

External boundary

The remote exact-SHA full qualification is intentionally deferred until the final release commit: no self-hosted runner is currently registered. Release and dry-run workflows fail closed until a fresh ephemeral runner produces the required same-run full evidence. This is one measured RTX 4080 stack, not continuous GPU CI; other hardware and distributed verl execution remain unmeasured. No scientific result or capability claim changed, and this PR does not tag or publish a release.

@DaoyuanLi2816
DaoyuanLi2816 marked this pull request as ready for review August 15, 2026 07:18
@DaoyuanLi2816
DaoyuanLi2816 merged commit 3301b8c into main Aug 16, 2026
16 checks passed
@DaoyuanLi2816
DaoyuanLi2816 deleted the v0.10.1-product-maturity branch August 16, 2026 06:24
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