DiLoCo: Distributed Low-Communication Training of Language Models

Arthur Douillard, Qixuan Feng, Andrei Alex Rusu, Rachita Chhaparia, Yani Donchev, Adhiguna Kuncoro, MarcAurelio Ranzato, Arthur Szlam, Jiajun Shen

Keywords: large-scale, language modeling, distributed learning, federated learning

TL;DR: Data-Parallel requires to all-reduce at every step, we propose to only do it only once every 500 steps and show good performance with LLM up to 400M scale.

Abstract: Large language models (LLM) have become a critical component in many applications of machine learning. However, standard approaches to training LLM require a large number of tightly interconnected accelerators, with devices exchanging gradients and other intermediate states at each optimization step. While it is difficult to build and maintain a single computing cluster hosting many accelerators, it might be easier to find several computing clusters each hosting a smaller number of devices. In this work, we propose a distributed optimization algorithm, Distributed Low-Communication (DiLoCo), that enables training of language models on islands of devices that are poorly connected. The approach is a variant of federated averaging, where the number of inner steps is large, the inner optimizer is AdamW, and the outer optimizer is Nesterov momentum. On the widely used C4 dataset, we show that DiLoCo on 8 workers performs as well as fully synchronous optimization while communicating 500 times less. DiLoCo exhibits great robustness to the data distribution of each worker. It is also robust to resources becoming unavailable over time, and vice versa, it can seamlessly leverage resources that become available during training.

Submission Number: 3