Repository navigation
Release v0.12.0: single-GPU verl v0.9 RL - #111
Merged
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
v0.9.0/483b8a009ba3a97563edee3a19887e4862b8094aFixes #89.
Compatibility boundary
PPO/GAE value mathematics are conformance-tested, but executable PPO remains classified
not_implemented: the runtime does not yet expose a separately trainable critic role, critic optimizer/checkpoint state or actor/reference/critic phase lifecycle. Ray, FSDP/FSDP2, Megatron and parallel degrees above one remaindistributed_onlyrather than being conflated with missing single-GPU semantics.The new GPU qualification is a runtime-correctness workload with a deterministic target-length formatting reward. It requires nonconstant within-group rewards, nonzero advantages, two parameter-changing Qwen3-0.6B NF4 LoRA updates and <=14.5 GiB peak reserved memory; it makes no task-quality claim.
Validation
Exact implementation commit
feed761514c3925dd2290d855baac6d38c417e50:ruff check .andruff format --check .mypy src/miniverlactionlint1.7.12pytest -q -m "not gpu and not network" --cov=miniverl: 2,566 passed, 10 skipped, 24 deselected, 83% branch coveragepytest -q -m gpu: 10 passed on one NVIDIA GeForce RTX 4080pytest -q -m network: 16 passedmkdocs build --strictand Playwright visual gate: 44 rendered SVG instances across 1440, 1024, 820 and 390 px viewports53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc; no benchmark file changedPublication remains gated on a new attempt-1, exact-release-SHA WSL2 RTX 4080 full qualification after merge, followed by the existing non-publishing dry run and OIDC workflow.