Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention

TL;DR

NSA introduces a hardware-aligned, trainable sparse attention mechanism that significantly speeds up long-context modeling in language models while maintaining or improving performance, enabling efficient training and inference.

Contribution

The paper presents NSA, a novel sparse attention method combining hierarchical strategies and hardware optimizations for efficient, end-to-end trainable long-context modeling.

Findings

NSA achieves substantial speedups on 64k sequences.

Models pretrained with NSA match or outperform full attention models.

NSA reduces pretraining computation without performance loss.

Abstract

Long-context modeling is crucial for next-generation language models, yet the high computational cost of standard attention mechanisms poses significant computational challenges. Sparse attention offers a promising direction for improving efficiency while maintaining model capabilities. We present NSA, a Natively trainable Sparse Attention mechanism that integrates algorithmic innovations with hardware-aligned optimizations to achieve efficient long-context modeling. NSA employs a dynamic hierarchical sparse strategy, combining coarse-grained token compression with fine-grained token selection to preserve both global context awareness and local precision. Our approach advances sparse attention design with two key innovations: (1) We achieve substantial speedups through arithmetic intensity-balanced algorithm design, with implementation optimizations for modern hardware. (2) We enable end-to-end training, reducing pretraining computation without sacrificing model performance. As shown in Figure 1, experiments show the model pretrained with NSA maintains or exceeds Full Attention models across general benchmarks, long-context tasks, and instruction-based reasoning. Meanwhile, NSA achieves substantial speedups over Full Attention on 64k-length sequences across decoding, forward propagation, and backward propagation, validating its efficiency throughout the model lifecycle.