SPLA: Block Sparse Plus Linear Attention for Long Context Modeling

TL;DR

SPLA introduces a block sparse plus linear attention framework that efficiently models long contexts by accurately selecting relevant blocks and compressing unselected ones, improving performance on long-context benchmarks.

Contribution

It proposes a novel selection metric based on second-order Taylor expansions and an optimized residual linear attention method to enhance long-context modeling efficiency.

Findings

Outperforms dense attention models on long-context benchmarks like RULER.

Maintains competitive general knowledge and reasoning capabilities.

Reduces IO overhead with an optimized subtraction-based RLA formulation.

Abstract

Block-wise sparse attention offers significant efficiency gains for long-context modeling, yet existing methods often suffer from low selection fidelity and cumulative contextual loss by completely discarding unselected blocks. To address these limitations, we introduce Sparse Plus Linear Attention (SPLA), a framework that utilizes a selection metric derived from second-order Taylor expansions to accurately identify relevant blocks for exact attention. Instead of discarding the remaining "long tail," SPLA compresses unselected blocks into a compact recurrent state via a residual linear attention (RLA) module. Crucially, to avoid IO overhead, we derive an optimized subtraction-based formulation for RLA -- calculating the residual as the difference between global and selected linear attention -- ensuring that unselected blocks are never explicitly accessed during inference. Our experiments demonstrate that SPLA closes the performance gap in continual pretraining, surpassing dense attention models on long-context benchmarks like RULER while maintaining competitive general knowledge and reasoning capabilities.