Communication-Efficient Language Model Training Scales Reliably and Robustly: Scaling Laws for DiLoCo
This addresses a critical bottleneck in scaling massive machine learning models for researchers and practitioners, though it is incremental as it builds on prior DiLoCo work.
The paper tackles the challenge of frequent synchronization slowdowns in data-parallel training for large language models by analyzing the scaling behavior of DiLoCo, finding that it scales predictably and robustly with model size, outperforming data-parallel training even at small sizes with benefits like increased optimal batch sizes and improved generalization.
As we scale to more massive machine learning models, the frequent synchronization demands inherent in data-parallel approaches create significant slowdowns, posing a critical challenge to further scaling. Recent work develops an approach (DiLoCo) that relaxes synchronization demands without compromising model quality. However, these works do not carefully analyze how DiLoCo's behavior changes with model size. In this work, we study the scaling law behavior of DiLoCo when training LLMs under a fixed compute budget. We focus on how algorithmic factors, including number of model replicas, hyperparameters, and token budget affect training in ways that can be accurately predicted via scaling laws. We find that DiLoCo scales both predictably and robustly with model size. When well-tuned, DiLoCo scales better than data-parallel training with model size, and can outperform data-parallel training even at small model sizes. Our results showcase a more general set of benefits of DiLoCo than previously documented, including increased optimal batch sizes, improved downstream generalization with scale, and improved evaluation loss for a fixed token budget.