📝 SummarXiv

Abstract

In large language models, the demand for modeling long contexts is ever-increasing, yet the quadratic complexity of standard self-attention presents a significant bottleneck. While existing sparse attention mechanisms enhance efficiency, they often suffer from limitations such as static patterns and information loss. This paper introduces a Trainable Dynamic Mask Sparse Attention mechanism that addresses these challenges through three key innovations. First, it leverages value vectors to dynamically generate content-aware sparse masks, enabling the model to adaptively identify and focus on crucial information. Second, it implements a position-aware sparse attention computation that effectively skips unnecessary computational regions. Finally, we ensure that the introduced dynamic masks and sparse weights do not obstruct gradients, thereby supporting end-to-end training. This dual-sparsity design allows the model to retain complete information while significantly reducing computational complexity, achieving an excellent balance between efficiency and performance. We validate the performance of Dynamic Mask Attention through comprehensive experiments. Comparative studies demonstrate that our method consistently achieves Pareto dominance across various tasks, including scaling laws, multi-query associative recall, general benchmarks, and needle-in-a-haystack tests, delivering up to 10 times acceleration. These results highlight its capability to effectively balance model efficiency with long-context modeling. Our computational kernel is open-sourced at https://github.com/SmallDoges/flash-dmattn to facilitate further research and application within the community.