DeMo: Decoupled Momentum Optimization

Bowen Peng, Jeffrey Quesnelle, Diederik P Kingma

Keywords: deep learning, large language models, optimization, training, generative models, pre-training, foundational models, distributed training

Abstract:

Training large scale neural networks typically involves sharing the gradients between all accelerators, which necessitates specialized high-speed interconnects. Taking cues from signal processing, we show that it is not necessary to share or synchronize the full optimizer states and model parameters during training. By decoupling the momentum and allowing divergence in the optimizer states across accelerators, it is possible to even improve convergence compared to previous state of the art optimizers.
From this, we introduce a Decoupled Momentum optimization algorithm (DeMo) that reduces the communication requirements by several orders of magnitude, potentially enabling future training of large neural networks on slow internet bandwidths with heterogeneous networking hardware. Furthermore, our method is agnostic to the network topology and neural network architecture, and supports scalable clock-synchronous distributed training with negligible compute and memory overhead.
Empirically, we show that models trained with DeMo match or surpass the performance of equal models trained with AdamW, entirely bypassing the need for high-speed interconnects for pre-training large scale foundation models.

Primary Area: foundation or frontier models, including LLMs

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5897