Reproduce “Accelerating Offloading-Based MoE Inference by Interleaving Concurrent Requests” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Hongseung Yu, In Gim, Kyunghan Lee, Lin Zhong

Bevy schedules concurrent requests in mixture-of-experts inference by letting each proceed when its missing experts arrive, revealing actual needs for the next layer. It batches ready requests to reuse shared weights while fetching known demands. With four or more requests and a 30% expert cache, it achieved 1.15–1.38× the fastest prefetching baseline’s decode throughput across tested models, workloads, and platforms.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.