LASP-2: Rethinking Sequence Parallelism for Linear Attention and Its Hybrid

TL;DR

LASP-2 introduces a novel sequence parallelism method that optimizes communication and computation for linear attention transformers, enabling efficient training on very long sequences with significant speed improvements.

Contribution

LASP-2 rethinks minimal communication in sequence parallelism for linear attention, reorganizing workflows to improve scalability and efficiency in distributed training.

Findings

Achieves 15.2% training speedup over LASP.

Attains 36.6% faster training than Ring Attention.

Supports training on sequences up to 2048K length across 64 GPUs.

Abstract

Linear sequence modeling approaches, such as linear attention, provide advantages like linear-time training and constant-memory inference over sequence lengths. However, existing sequence parallelism (SP) methods are either not optimized for the right-product-first feature of linear attention or use a ring-style communication strategy, which results in lower computation parallelism, limits their scalability for longer sequences in distributed systems. In this paper, we introduce LASP-2, a new SP method to enhance both communication and computation parallelism when training linear attention transformer models with very-long input sequences. Compared to previous work LASP, LASP-2 rethinks the minimal communication requirement for SP on linear attention layers, reorganizes the whole communication-computation workflow of LASP. In this way, only one single AllGather collective communication is needed on intermediate memory states, whose sizes are independent of the sequence length, leading to significant improvements of both communication and computation parallelism, as well as their overlap. Additionally, we extend LASP-2 to LASP-2H by applying similar communication redesign to standard attention modules, offering an efficient SP solution for hybrid models that blend linear and standard attention layers. Our evaluation on a Linear-Llama3 model, a variant of Llama3 with linear attention replacing standard attention, demonstrates the effectiveness of LASP-2 and LASP-2H. Specifically, LASP-2 achieves training speed improvements of 15.2% over LASP and 36.6% over Ring Attention, with a sequence length of 2048K across 64 GPUs. The Code is released as a part of: https://github.com/OpenSparseLLMs/Linear-MoE.