[rl] Refactor rollout decoding across vLLM and Levanter#5035
Open
taivu1998 wants to merge 6 commits intomarin-community:mainfrom
Open
[rl] Refactor rollout decoding across vLLM and Levanter#5035taivu1998 wants to merge 6 commits intomarin-community:mainfrom
taivu1998 wants to merge 6 commits intomarin-community:mainfrom
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: fd0345d2ea
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Unify RL rollout decoding behind a shared DecodingConfig and record applied decoding on rollouts so generation behavior stays explicit and reproducible. Expand the vLLM path to honor the richer decode surface, and make the Levanter wrapper fail loudly when a requested knob is not actually supported.
Make the Levanter RL inference wrapper explicit about the decoding fields it actually honors today. This keeps the shared decoding contract honest while preserving stop-token fallback behavior through focused regression coverage.
Move train decoding defaults into curriculum construction and leave the vLLM sampling config as backend fallback only. This keeps lesson config as the clear runtime source of truth and avoids implying that builder fields control live rollout counts or top-k behavior.
Separate vLLM runtime settings from fallback sampling defaults so rollout policy stays distinct from engine tuning. This keeps future vLLM performance knobs from re-mixing with decode semantics after the PR1-PR4 cleanup.
Carry top_p through the Levanter inference request path, decode state, and sampler so Marin RL can use nucleus sampling on the native backend. This keeps the Levanter wrapper honest while closing the highest-value remaining decode-surface gap after the earlier decoding cleanup.
taivu1998
force-pushed
the
tdv/rl-pr1-shared-decoding
branch
from
9824578 to
50fdbe2
Compare
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
1 participant
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.
Unify RL rollout decoding around a shared configuration, record applied decoding on rollouts, and keep curriculum-owned rollout policy separate from backend fallbacks. Expand the vLLM surface, tighten Levanter validation, and add real Levanter top_p support through the native inference engine. Validated with the non-slow RL suite and focused Levanter inference tests.