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

As far as I know, SGLang manages its own KV pages, and has no spill-to-remote-tier, like cloud storage, which NVIDIA KV Block manager and NIXL enable. I am leaning towards the first implementation below. Would love to hear how the SGLang team thinks!

Option 1 – on-demand pages. When SGLang asks for a KV page, we call the KVBM adapter, wrap that page in a KVBM Block, and let KVBM persist it to NIXL. Same as vllm's mega PR

Option 2 – pre-alloc. Have KVBM hand SGLang a single 20 GB “free_pages” region up front.

Related resources

Implmenetation on the nvidia side: ai-dynamo/dynamo#1199