In plain words: Instead of linking every computer into one tightly connected cluster, each small group trains alone for many steps and only swaps updates with the others occasionally. Across 8 groups it matched normal training that shares every step, while sending 500 times less information.
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.
Arthur Douillard, Qixuan Feng, Andrei A. Rusu, Rachita Chhaparia, Yani Donchev, Adhiguna Kuncoro, Marc'Aurelio Ranzato, Arthur Szlam, Jiajun Shen
arXiv:2311.08105 · cs.LG, cs.CL · submitted Nov 14, 2023 · updated Sep 23, 2024
abstract · pdf · html
That said, there is a fundamental rule here -- gradients must, at some point, be shared. This paper says they don't have to be shared every step. Which is really great. But, gradient sizes are on the order of magnitude of the network being trained.
To get this useful for indie LLM training, I'd guess that you'll want it to work sharing gradients not more often than once every few hours to day -- more than that, and bandwidth costs sending 50GB weights files around during training are going to kill you.
The paper seems to indicate they can share every few thousand? (8 workers x 500 times bandwidth reduction?) steps, but sharing every few millions would help that distributed large LLM use case, and my guess is that this is going to be a bit harder to get to converge. But hopefully not!
Related: there's an n^2 problem for getting the gradients out to lots of distributed compute resources.
So, if anyone wants to work on step 2, I'd suggest there's an engineering task: 1-N gradient upload and distribution, and then a science task: revalidating this with a few orders of magnitude less common checkins.