Skip to content

Conversation

finbarrtimbers
Copy link
Collaborator

vllm_utils.py was not used anywhere.

finbarrtimbers and others added 28 commits October 8, 2025 10:28
* uses uni

* undid multipprocessing

* Cleaned up logging (#1069)

* Fix missing logprobs accumulation in tool-using requests

When tools are enabled, process_request_async accumulates tokens across
multiple tool-calling iterations but was discarding logprobs by setting
them to None. This caused crashes when the downstream code tried to
extract logprobs from the CompletionOutput.

Fixed by:
- Adding accumulated_logprobs list to track logprobs across iterations
- Accumulating output.logprobs for model-generated tokens
- Creating mock logprob objects with NaN values for tool output tokens
- Passing accumulated_logprobs to CompletionOutput instead of None

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

* Fixed tool trigger check

* Updated code

---------

Co-authored-by: Claude <[email protected]>
@finbarrtimbers
Copy link
Collaborator Author

Blocked by #1016.

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