SMILE: Scaling Mixture-of-Experts with Efficient Bi-level Routing
This addresses a scalability bottleneck for researchers and practitioners using large-scale MoE models, representing an incremental improvement in routing efficiency.
The paper tackles the training efficiency degradation in mixture-of-experts (MoE) models as the number of experts increases, caused by network congestion in routing, and introduces SMILE with bi-level routing to achieve a 2.5x speedup in pretraining throughput over Switch Transformer without convergence loss.
The mixture of Expert (MoE) parallelism is a recent advancement that scales up the model size with constant computational cost. MoE selects different sets of parameters (i.e., experts) for each incoming token, resulting in a sparsely-activated model. Despite several successful applications of MoE, its training efficiency degrades significantly as the number of experts increases. The routing stage in MoE relies on the efficiency of the All2All communication collective, which suffers from network congestion and has poor scalability. To mitigate these issues, we introduce SMILE, which exploits heterogeneous network bandwidth and splits a single-step routing into bi-level routing. Our experimental results show that the proposed method obtains a 2.5x speedup over Switch Transformer in terms of pretraining throughput on the Colossal Clean Crawled Corpus without losing any convergence speed.