Janus Moe Disaggregation

Enable scalable MoE inference by disaggregating attention and expert layers onto independent GPU sub-clusters. Use adaptive two-phase communication, activation load-balanced scheduling, and activation-aware expert management. Achieve 3.9× higher per-GPU throughput than state-of-the-art systems.

adu2021 Updated 7 repo stars

File contents

adu2021/skillxiv commit 03aebb2853

Frequently asked questions

npx get-skill add adu2021/janus-moe-disaggregation