Abstract: Training Mixture-of-Experts (MoE) models introduces sparse and highly imbalanced all-to-all communication that dominates iteration time. Conventional load-balancing methods fail to exploit ...
Abstract: Distributed training is the most common way to scale out and accelerate Deep Neural Network (DNN) training. Distributed DNN training requires synchronization of gradient aggregation among ...