ST-MoE: Designing Stable and Transferable Sparse Expert Models

TL;DR

This paper introduces ST-MoE, a scalable and stable sparse expert model that achieves state-of-the-art transfer learning performance across diverse NLP tasks while maintaining computational efficiency.

Contribution

It provides a design guide to address training instability in MoE models and demonstrates scaling a sparse model to 269B parameters with high transferability.

Findings

Achieved state-of-the-art transfer performance on multiple NLP benchmarks.

Scaled a sparse model to 269B parameters with efficiency comparable to smaller dense models.

Addressed training stability issues in large sparse models.

Abstract

Scale has opened new frontiers in natural language processing -- but at a high cost. In response, Mixture-of-Experts (MoE) and Switch Transformers have been proposed as an energy efficient path to even larger and more capable language models. But advancing the state-of-the-art across a broad set of natural language tasks has been hindered by training instabilities and uncertain quality during fine-tuning. Our work focuses on these issues and acts as a design guide. We conclude by scaling a sparse model to 269B parameters, with a computational cost comparable to a 32B dense encoder-decoder Transformer (Stable and Transferable Mixture-of-Experts or ST-MoE-32B). For the first time, a sparse model achieves state-of-the-art performance in transfer learning, across a diverse set of tasks including reasoning (SuperGLUE, ARC Easy, ARC Challenge), summarization (XSum, CNN-DM), closed book question answering (WebQA, Natural Questions), and adversarially constructed tasks (Winogrande, ANLI R3).