Repository navigation
fix: make logprobs capture robust to top_logprobs=null in vllm model - #1612
Merged
ananthsub merged 1 commit intoJun 16, 2026
Merged
Conversation
vLLM computes `logprobs = top_logprobs if logprobs else None`, so a request carrying an explicit top_logprobs=null (common in replayed rollout JSONL) makes vLLM return no logprobs. That empties the captured token ids and silently zeroes the training loss mask, which freezes RL training with no error. - Pin top_logprobs=0 when return_token_id_information is set, so capture is independent of any value inherited from the request or dataset. - Drop a null top_logprobs on the non-capture path before forwarding. - Raise a clear error when capture is requested but vLLM returns no logprobs, instead of an opaque TypeError or silently empty training data. Add unit tests covering all three. Signed-off-by: Ananth Subramaniam <ansubramania@nvidia.com>
This was referenced Jun 16, 2026
cmunley1
approved these changes
Jun 16, 2026
ritaneves
pushed a commit
that referenced
this pull request
Jun 25, 2026
…1612) vLLM computes `logprobs = top_logprobs if logprobs else None`, so a request carrying an explicit `top_logprobs` field makes vLLM return no logprobs. That empties the captured token ids and silently zeroes the training loss mask, which freezes RL training with no error. - Pin `top_logprobs=0` when `return_token_id_information` is set, so capture is independent of any value inherited from the request or dataset - Drop a null `top_logprobs` on the non-capture path before forwarding - Raise a clear error when capture is requested but vLLM returns no logprobs, instead of an opaque TypeError or silently empty training data supersedes #1013 Signed-off-by: Ananth Subramaniam <ansubramania@nvidia.com> Signed-off-by: Rita Fernandes Neves <rfernandesne@nvidia.com>
This branch was previously deployed
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.
vLLM computes
logprobs = top_logprobs if logprobs else None, so a request carrying an explicittop_logprobsfield makes vLLM return no logprobs. That empties the captured token ids and silently zeroes the training loss mask, which freezes RL training with no error.top_logprobs=0whenreturn_token_id_informationis set, so capture is independent of any value inherited from the request or datasettop_logprobson the non-capture path before forwardingsupersedes #1013