A Distributed Data-Parallel PyTorch Implementation of the Distributed Shampoo Optimizer for Training Neural Networks At-Scale

TL;DR

This paper presents a distributed, data-parallel implementation of the Shampoo optimizer in PyTorch, enabling scalable training of neural networks with minimal performance overhead and demonstrating its effectiveness on ImageNet ResNet50.

Contribution

The work introduces a complete, optimized PyTorch implementation of Shampoo that supports multi-GPU distributed training and demonstrates its advantages over standard methods.

Findings

Achieves at most 10% slower per-step time compared to diagonal methods.

Demonstrates Shampoo's superiority on ImageNet ResNet50 with minimal hyperparameter tuning.

Enables scalable training of deep networks at scale.

Abstract

Shampoo is an online and stochastic optimization algorithm belonging to the AdaGrad family of methods for training neural networks. It constructs a block-diagonal preconditioner where each block consists of a coarse Kronecker product approximation to full-matrix AdaGrad for each parameter of the neural network. In this work, we provide a complete description of the algorithm as well as the performance optimizations that our implementation leverages to train deep networks at-scale in PyTorch. Our implementation enables fast multi-GPU distributed data-parallel training by distributing the memory and computation associated with blocks of each parameter via PyTorch's DTensor data structure and performing an AllGather primitive on the computed search directions at each iteration. This major performance enhancement enables us to achieve at most a 10% performance reduction in per-step wall-clock time compared against standard diagonal-scaling-based adaptive gradient methods. We validate our implementation by performing an ablation study on training ImageNet ResNet50, demonstrating Shampoo's superiority over standard training recipes with minimal hyperparameter tuning.