Lancet: Accelerating Mixture-of-Experts Training via Whole Graph Computation-Communication Overlapping

TL;DR

This paper introduces Lancet, a system that enhances Mixture-of-Experts training by overlapping communication and computation at the graph level, significantly reducing communication time and improving training speed.

Contribution

Lancet extends overlap techniques to the entire training graph using compiler-based optimization, achieving substantial communication reduction and speedup.

Findings

Reduces communication time by up to 77%.

Achieves up to 1.3x end-to-end training speedup.

Effectively overlaps non-MoE and gradient computations.

Abstract

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.