MARS: Unleashing the Power of Variance Reduction for Training Large Models

TL;DR

This paper introduces MARS, a unified framework that combines variance reduction with preconditioned gradient methods, significantly improving the efficiency of training large neural models like GPT-2.

Contribution

The paper proposes MARS, a novel optimization framework that integrates variance reduction with preconditioned gradient methods for large-scale neural network training.

Findings

MARS outperforms AdamW in training GPT-2 models.

Three instances of MARS leverage AdamW, Lion, and Shampoo.

Experimental results show large margin improvements over existing optimizers.

Abstract

Training deep neural networks--and more recently, large models demands efficient and scalable optimizers. Adaptive gradient algorithms like Adam, AdamW, and their variants have been central to this task. Despite the development of numerous variance reduction algorithms in the past decade aimed at accelerating stochastic optimization in both convex and nonconvex settings, variance reduction has not found widespread success in training deep neural networks or large language models. Consequently, it has remained a less favored approach in modern AI. In this paper, to unleash the power of variance reduction for efficient training of large models, we propose a unified optimization framework, MARS (Make vAriance Reduction Shine), which reconciles preconditioned gradient methods with variance reduction via a scaled stochastic recursive momentum technique. Within our framework, we introduce three instances of MARS that leverage preconditioned gradient updates based on AdamW, Lion, and Shampoo, respectively. We also draw a connection between our algorithms and existing optimizers. Experimental results on training GPT-2 models indicate that MARS consistently outperforms AdamW by a large margin. The implementation of MARS is available at https://github.com/AGI-Arena/MARS.