SimpleFSDP: Simpler Fully Sharded Data Parallel with torch.compile

TL;DR

SimpleFSDP introduces a PyTorch-native, compiler-friendly Fully Sharded Data Parallel framework that simplifies implementation, enhances performance through compiler optimizations, and reduces memory usage in large model training.

Contribution

It presents a novel, torch.compile-friendly FSDP implementation with IR node bucketing and reordering for better computation-communication overlap, improving distributed training efficiency.

Findings

Up to 28.54% memory reduction in large models.

68.67% throughput improvement over existing FSDP2.

Effective use of compiler optimizations for distributed training.

Abstract

Distributed training of large models consumes enormous computation resources and requires substantial engineering efforts to compose various training techniques. This paper presents SimpleFSDP, a PyTorch-native compiler-based Fully Sharded Data Parallel (FSDP) framework, which has a simple implementation for maintenance and composability, allows full computation-communication graph tracing, and brings performance enhancement via compiler backend optimizations. SimpleFSDP's novelty lies in its unique $torch.compile$-friendly implementation of collective communications using existing PyTorch primitives, namely parametrizations, selective activation checkpointing, and DTensor. It also features the first-of-its-kind intermediate representation (IR) nodes bucketing and reordering in the TorchInductor backend for effective computation-communication overlapping. As a result, users can employ the aforementioned optimizations to automatically or manually wrap model components for minimal communication exposure. Extensive evaluations of SimpleFSDP on Llama 3 models (including the ultra-large 405B) using TorchTitan demonstrate up to 28.54% memory reduction and 68.67% throughput improvement compared to the most widely adopted FSDP2 eager framework, when composed with other distributed training techniques.