Optimizing Native Sparse Attention with Latent Attention and Local Global Alternating Strategies

TL;DR

This paper enhances Native Sparse Attention by introducing alternating local and global strategies and Latent Attention mechanisms, significantly improving long-sequence modeling while reducing memory usage.

Contribution

It proposes a novel alternating attention strategy combined with Latent Attention to improve long-context modeling and efficiency in sparse attention methods.

Findings

Reduces KV-cache memory by 50% compared to NSA.

Matches or exceeds full attention in reasoning and long-text tasks.

Effective across models from 340M to 1.3B parameters.

Abstract

In this work, we conduct a systematic analysis of Native Sparse Attention (NSA) and propose targeted improvements that enhance long-context modeling. A key insight is that alternating between local (sliding-window) and global (compression, selective) attention across layers, rather than using fixed patterns, enables more effective propagation of long-range dependencies and substantially boosts performance on long-sequence tasks. Meanwhile, we further refine NSA's branches with Latent Attention that the sliding-window branch is enhanced with Multi-head Latent Attention (MLA) while compression and selective branches adopt Group-head Latent Attention (GLA). These changes reduce KV-cache memory by 50\% versus NSA while improving the model's common-sense reasoning and long-text understanding capabilities. Experiments on models from 340M to 1.3B parameters (trained on 15B and 100B tokens) show our method matches or exceeds full attention and native sparse attention in both common-sense reasoning and long-context understanding tasks.