1 paper
Lizhang Chen, Jonathan Li, Qi Wang +5
Mixture-of-Experts (MoE) models rely on balanced expert utilization to fully realize their scalability. However, existing load-balancing methods are largely heuristic and operate o…