[Feature][NVIDIA] EPLB enablement for DSR1 disagg on GB200

[wenscarl]

Checklist

• If this is not a feature request but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
• Please use English. Otherwise, it will be closed.

Motivation

This issue tracks the ongoing work originating from NVIDIA.

1. The current EPLB implementation is broken because the first dimension of the global scaling factor does not match num_local_expert. As a temporary workaround, see Fix EPLB + FP4 Quantization Compatibility Issue #13715. @shifangx
   @wenscarl
2. NVIDIA has proposed a Linear Programming (LP)–based expert-parallelism algorithm. The integration will proceed in two stages:

a) Integrate the LP kernel into FlashInfer. @feliang-git

b) Integrate the FlashInfer operator into sglang.

Related resources

No response

[github-actions]

This issue has been automatically closed due to inactivity. Please feel free to reopen it if needed.