Rectified Sparse Attention

TL;DR

Rectified Sparse Attention (ReSA) enhances long-sequence generation efficiency in large language models by combining sparse attention with periodic dense cache rectification, significantly improving speed while maintaining quality.

Contribution

ReSA introduces a novel method that bounds error accumulation in sparse attention by periodic dense cache refreshes, improving long-sequence generation quality and efficiency.

Findings

Achieves up to 2.42× speedup at 256K sequence length.

Maintains near-lossless generation quality across tasks.

Demonstrates effectiveness in math reasoning, language modeling, and retrieval.

Abstract

Efficient long-sequence generation is a critical challenge for Large Language Models. While recent sparse decoding methods improve efficiency, they suffer from KV cache misalignment, where approximation errors accumulate and degrade generation quality. In this work, we propose Rectified Sparse Attention (ReSA), a simple yet effective method that combines block-sparse attention with periodic dense rectification. By refreshing the KV cache at fixed intervals using a dense forward pass, ReSA bounds error accumulation and preserves alignment with the pretraining distribution. Experiments across math reasoning, language modeling, and retrieval tasks demonstrate that ReSA achieves near-lossless generation quality with significantly improved efficiency. Notably, ReSA delivers up to 2.42$\times$ end-to-end speedup under decoding at 256K sequence length, making it a practical solution for scalable long-context inference. Code is available at https://aka.ms/ReSA-LM.