Skip to content

Conversation

@EduardDurech
Copy link
Collaborator

Support SGLang remote_instance as well as different {rank0,k-ary,epidemic} seed strategies

Requires sgl-project/sglang#16410

Copy link
Contributor

@gemini-code-assist gemini-code-assist bot left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review

This pull request introduces support for SGLang's remote_instance feature, enabling more efficient distributed model weight loading. It adds a RemoteInstanceSeedCoordinator Ray actor to manage seed replicas and supports various seeding strategies (rank0, k-ary, epidemic). The changes are well-structured, including new port management utilities and updates to the server launch logic. The implementation correctly uses a lease-like mechanism for seed servers, with a finally block to ensure they are returned. However, I've identified a high-severity issue in the error handling within this finally block, which could lead to resource leaks and potential deadlocks.

@EduardDurech EduardDurech marked this pull request as draft January 11, 2026 23:37
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