Rimuru: Unified Memory Management for Modern LLM Serving via Elastic Expert Offloading
Unified page pool shares memory between model weights and attention states.
Explainer · on the wire
The explainer for this story is still on the wire.
The claim, the mechanism, and the figures land here once generated.
Source · arXiv