FlexDeMo: Decoupled Momentum Optimization for Hybrid Sharded Data Parallel Training

Training large neural network models requires extensive computational resources, often distributed across several nodes and accelerators. Recent findings suggest that it may be sufficient to only exchange the fast moving components of the gradients, while accumulating momentum locally (Decoupled Momentum, or DeMo). However, when considering larger models that do not fit on a single accelerator, the exchange of gradient information and the integration of DeMo needs to be reconsidered. Here, we propose employing a hybrid sharded data parallel training strategy, FlexDeMo, whereby nodes fully shard model parameters locally between different accelerators, while inter-node communication bandwidth requirements are reduced by synchronizing only fast-moving components instead of the full gradients. This effectively combines previous hybrid sharded strategies with the advantages of decoupled momentum. Our experimental results show that FlexDeMo is on par with hybrid sharded data parallel training employing AdamW and full gradient synchronization in terms of validation loss, demonstrating its viability. Furthermore, FlexDeMo achieves improved training speed compared to full gradient synchronization across nodes. In a bandwidth-constrained 2-node setup, FlexDeMo allows reaching desired levels of validation loss faster than hybrid sharded data parallel training with full gradient synchronization.