vllm-project
[RFC]: Implement Structured Output support for V1 engine
Motivation.
Structured Output is supported in v0, but not yet in v1. One reason for the delay is there have been performance challenges with the integration in v0, and we'd like to rethink the integration approach. We would also like to account for supporting additional techniques, jump decoding in particular, in the future.
The document below covers the proposed integration of the Structured Output functionality in V1 of the vLLM engine.
Proposed Change.
A draft proposal can be found in this google doc: https://docs.google.com/document/d/1H6m_Y3FLJ1FYGCmjXdZzoJv-JCDSxnKuSY2XiAj-c6c/edit?tab=t.0
This content will eventually be moved into a PR as an addition to the design docs section of the vllm docs.
Related issue for closing xgrammar feature gaps: https://github.com/vllm-project/vllm/issues/12131
Feedback Period.
No response
CC List.
@mgoin @aarnphm @markmc @simon-mo @xuechendi @WoosukKwon
Any Other Things.
No response
Before submitting a new issue...
- [X] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.
