good first issue
Repository metrics
- Stars
- (28,442 stars)
- PR merge metrics
- (Avg merge 2d 1h) (1,000 merged PRs in 30d)
Description
Checklist
- 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
- 2. Please use English, otherwise it will be closed.
Motivation
To support a flexible and reliable load balancing strategy, external router (sgl-router) and internal controller (dp-controller) should be aware of the real-time load situation.
Current implementation uses a polling mechanism and a watch-based communicator to query load from the scheduler. However, this is flexible and elegant.
- Implement a piggyback load reporting mechanism, in which load information is carried through restreaming token messages.
- Also keep the watching mode, always storing the latest load information in http worker for external query.
Related resources
No response