Scaling Laws and Compute-Optimal Training Beyond Fixed Training Durations

TL;DR

This paper explores alternative training schedules to cosine decay, demonstrating predictable scaling behavior and the benefits of stochastic weight averaging, enabling more efficient scaling experiments with reduced compute resources.

Contribution

It introduces a simple, predictable training schedule alternative and shows how to perform scalable experiments efficiently, improving understanding of model scaling beyond fixed training durations.

Findings

Constant learning rate with cooldowns scales predictably.

Stochastic weight averaging improves performance during training.

Scaling experiments can be done with fewer runs and less compute.

Abstract

Scale has become a main ingredient in obtaining strong machine learning models. As a result, understanding a model's scaling properties is key to effectively designing both the right training setup as well as future generations of architectures. In this work, we argue that scale and training research has been needlessly complex due to reliance on the cosine schedule, which prevents training across different lengths for the same model size. We investigate the training behavior of a direct alternative -- constant learning rate and cooldowns -- and find that it scales predictably and reliably similar to cosine. Additionally, we show that stochastic weight averaging yields improved performance along the training trajectory, without additional training costs, across different scales. Importantly, with these findings we demonstrate that scaling experiments can be performed with significantly reduced compute and GPU hours by utilizing fewer but reusable training runs. Our code is available at \url{https://github.com/epfml/schedules-and-scaling/}.