Untied Ulysses: Memory-Efficient Context Parallelism via Headwise Chunking

TL;DR

This paper introduces UPipe, a memory-efficient context parallelism method that enables training of Transformer models with much longer sequences by reducing activation memory usage through headwise chunking.

Contribution

UPipe is a novel technique that performs fine-grained chunking at the attention head level to significantly reduce memory usage and support longer context lengths in Transformer training.

Findings

Reduces activation memory by up to 87.5% for 32B Transformers.

Supports context length of 5 million tokens for Llama3-8B on a single node.

Matches previous methods in training speed while enabling longer sequences.

Abstract

Efficiently processing long sequences with Transformer models usually requires splitting the computations across accelerators via context parallelism. The dominant approaches in this family of methods, such as Ring Attention or DeepSpeed Ulysses, enable scaling over the context dimension but do not focus on memory efficiency, which limits the sequence lengths they can support. More advanced techniques, such as Fully Pipelined Distributed Transformer or activation offloading, can further extend the possible context length at the cost of training throughput. In this paper, we present UPipe, a simple yet effective context parallelism technique that performs fine-grained chunking at the attention head level. This technique significantly reduces the activation memory usage of self-attention, breaking the activation memory barrier and unlocking much longer context lengths. Our approach reduces intermediate tensor memory usage in the attention layer by as much as 87.5$\%$ for 32B Transformers, while matching previous context parallelism techniques in terms of training speed. UPipe can support the context length of 5M tokens when training Llama3-8B on a single 8$\times$H100 node, improving upon prior methods by over 25$\%$.