Skip to content

Keep policy versions unique when a trainer resumes behind the store's releases - #328

Merged
michaellchung merged 1 commit into
mainfrom
resume-version-account
Nov 3, 2026
Merged

michaellchung merged 1 commit into
mainfrom
resume-version-account

Conversation

@michaellchung

Copy link
Copy Markdown
Contributor

Summary

  • bridges: a buffer trainer whose round cut is behind the store's newest release continues the version account above it (resume_at with behind_round_cuts; LocalOnlySync.start), so a resumed run never re-publishes a version that already exists.
  • Miles single trainer: on relaunch the publisher continues above the store's releases (_resume_single_trainer, ReleasingPublisher resumed_base) instead of reusing a version already released.

Behavior changes

  • Miles single trainer, intentional: after a relaunch, LATEST in the release store no longer moves back to the resumed checkpoint's version. The first release of the relaunched trainer gets the next version above the store's newest release, so versions stay unique and monotonic; samplers that follow LATEST never see it regress.
  • Buffer trainers resuming from a cut behind the store's releases number their releases the same way.

Test plan

  • pytest tests/test_rl_sampler_island.py tests/test_rl_verl_buffer_trainer.py plus the earlier stacked suites: 290 passed.
  • scripts/precommit_check.py on the changed files: no errors.

Stacked on #327.

🤖 Generated with Claude Code

@michaellchung
michaellchung force-pushed the background-release-mfu branch from 4dfe1a1 to d564b00 Compare November 2, 2026 17:31
@michaellchung
michaellchung force-pushed the resume-version-account branch from 9fce8ce to d770015 Compare November 2, 2026 17:31
@michaellchung
michaellchung force-pushed the background-release-mfu branch from d564b00 to ec4c078 Compare November 2, 2026 17:42
@michaellchung
michaellchung force-pushed the resume-version-account branch from d770015 to e8a6c81 Compare November 2, 2026 17:42
@michaellchung
michaellchung force-pushed the background-release-mfu branch from ec4c078 to fa54896 Compare November 2, 2026 18:30
@michaellchung
michaellchung force-pushed the resume-version-account branch from e8a6c81 to 2aed9e1 Compare November 2, 2026 18:30
@michaellchung
michaellchung force-pushed the background-release-mfu branch from fa54896 to 91226de Compare November 2, 2026 18:42
@michaellchung
michaellchung force-pushed the resume-version-account branch from 2aed9e1 to 1caf025 Compare November 2, 2026 18:42
@michaellchung
michaellchung force-pushed the background-release-mfu branch from 91226de to 7ed7c17 Compare November 2, 2026 19:27
@michaellchung
michaellchung force-pushed the resume-version-account branch from 1caf025 to 00c334d Compare November 2, 2026 19:27
@michaellchung
michaellchung force-pushed the background-release-mfu branch from 7ed7c17 to c60a94d Compare November 2, 2026 19:58
@michaellchung
michaellchung force-pushed the resume-version-account branch 2 times, most recently from bf73f0c to dc0309e Compare November 2, 2026 20:09
@michaellchung
michaellchung force-pushed the background-release-mfu branch from c60a94d to 933f436 Compare November 2, 2026 20:09
@michaellchung
michaellchung force-pushed the background-release-mfu branch from 933f436 to 7ec663b Compare November 3, 2026 02:05
@michaellchung
michaellchung force-pushed the resume-version-account branch from dc0309e to a9212e7 Compare November 3, 2026 02:05
@michaellchung
michaellchung force-pushed the background-release-mfu branch from 7ec663b to 6ef9f6e Compare November 3, 2026 03:29
@michaellchung
michaellchung force-pushed the resume-version-account branch from a9212e7 to f424f86 Compare November 3, 2026 03:29
@michaellchung
michaellchung changed the base branch from background-release-mfu to main November 3, 2026 04:08
… releases

- bridges: a buffer trainer whose round cut is behind the store's newest
  release continues the version account above it (resume_at,
  behind_round_cuts; LocalOnlySync.start).
- Miles single trainer: on relaunch the publisher continues above the
  store's releases (_resume_single_trainer, ReleasingPublisher resumed_base)
  instead of reusing a version already released.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@michaellchung
michaellchung force-pushed the resume-version-account branch from f424f86 to 38eee6a Compare November 3, 2026 04:09
@michaellchung
michaellchung merged commit 78dedcd into main Nov 3, 2026
2 checks passed
@michaellchung
michaellchung deleted the resume-version-account branch November 3, 2026 04:53
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