Untied Ulysses: Memory-Efficient Context Parallelism via Headwise Chunking
This work addresses memory limitations in training long-sequence Transformers, enabling longer contexts with minimal speed loss, though it is incremental as it builds on existing parallelism techniques.
The paper tackles the memory inefficiency of existing context parallelism methods for Transformer models by introducing UPipe, a technique that performs fine-grained chunking at the attention head level, reducing activation memory usage by up to 87.5% for 32B Transformers and supporting context lengths of 5M tokens for Llama3-8B on a single node, a 25% improvement over prior methods.
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$\%$.