Lancet: Accelerating Mixture-of-Experts Training via Whole Graph Computation-Communication Overlapping
This addresses a bottleneck in training large-scale MoE models for AI applications, offering incremental improvements over existing overlapping methods.
The paper tackles the problem of extended all-to-all communication latency in Mixture-of-Experts (MoE) training by proposing Lancet, a system that achieves whole graph computation-communication overlapping, reducing non-overlapping communication time by up to 77% and achieving an end-to-end speedup of up to 1.3× compared to state-of-the-art methods.
The Mixture-of-Expert (MoE) technique plays a crucial role in expanding the size of DNN model parameters. However, it faces the challenge of extended all-to-all communication latency during the training process. Existing methods attempt to mitigate this issue by overlapping all-to-all with expert computation. Yet, these methods frequently fall short of achieving sufficient overlap, consequently restricting the potential for performance enhancements. In our study, we extend the scope of this challenge by considering overlap at the broader training graph level. During the forward pass, we enable non-MoE computations to overlap with all-to-all through careful partitioning and pipelining. In the backward pass, we achieve overlap with all-to-all by scheduling gradient weight computations. We implement these techniques in Lancet, a system using compiler-based optimization to automatically enhance MoE model training. Our extensive evaluation reveals that Lancet significantly reduces the time devoted to non-overlapping communication, by as much as 77%. Moreover, it achieves a notable end-to-end speedup of up to 1.3 times when compared to the state-of-the-art solutions.