Trainable Dynamic Mask Sparse Attention

TL;DR

This paper introduces a trainable dynamic mask sparse attention mechanism that adaptively combines position-aware and content-aware strategies, significantly improving efficiency and performance in long-context modeling for large language models.

Contribution

It proposes a novel dynamic mask sparse attention method that is fully differentiable, adaptable, and hardware-efficient, advancing sparse attention techniques for large language models.

Findings

Consistently outperforms state-of-the-art sparse attention baselines.

Achieves up to 10x speedup in computational efficiency.

Effectively balances model efficiency with long-context modeling capabilities.

Abstract

The increasing demand for long-context modeling in large language models (LLMs) is bottlenecked by the quadratic complexity of the standard self-attention mechanism. The community has proposed sparse attention to mitigate this issue. However, position-aware sparse attention methods rely on static sparse structures that lack adaptability to diverse query contexts, while content-aware sparse attention methods depend on heuristic key-value selection, hindering full differentiability. We introduce a trainable dynamic mask sparse attention mechanism, a method that merges the advantages of both position-aware and content-aware approaches. Dynamic Mask Attention (DMA) achieves this through three key innovations: First, it leverages value vector representations to generate content-aware dynamic masks, enabling the model to adaptively identify and attend to critical information. Second, it computes position-aware sparse weights in a hardware-friendly manner, efficiently skipping unnecessary computational regions. Finally, we demonstrate that the introduced dynamic mask and sparse weights do not obstruct gradients, supporting end-to-end training. We have validated the performance of DMA through comprehensive experiments. A large body of experimental evidence shows that DMA consistently holds a Pareto advantage over state-of-the-art sparse attention baselines in tasks including scaling laws, multi-query associative recall, standard benchmarks, and needle in a haystack tests, while also delivering up to a 10x overall speedup. These results highlight its ability to effectively balance model efficiency with long-context modeling capabilities. Our computational kernel code is now open-source at https://github.com/SmallDoges/flash-dmattn to encourage further research and application by the community.