Scaling neural network training increasingly depends on synchronous data-parallelism, yet full-precision gradient all-reduce imposes a severe communication bottleneck. We propose Decoupled Momentum Optimization (DeMo), a drop-in replacement for any momentum-based optimizers that significantly reduces the communication bandwidth while maintaining convergence. DeMo (i) decouples local momentum updates, (ii) applies a fast orthonormal transform (e.g., DCT) followed by top-k sparsification, and (iii) reuses the momentum buffer as error feedback via momentum subtraction. This design reduces per-step communication by up to two orders of magnitude with minimal computational overhead. Experiments on 300M and 1B-parameter DeMo language models show DeMo transmits up to 85x less data per GPU than AdamW-DDP while achieving comparable loss and accuracy. DeMo is topology-agnostic and enables training across multi-datacenter or Ethernet-based setups. Code is available at https://github.com/bloc97/DeMo
@article{arxiv.2411.19870,
title = {DeMo: Decoupled Momentum Optimization},
author = {Bowen Peng and Lizhang Chen and Baiyu Su and Jeffrey Quesnelle and Diederik P. Kingma and Qiang Liu},
journal= {arXiv preprint arXiv:2411.19870},
year = {2026}
}