Structured sparse attention for accelerating long-context diffusion language models
Abstract
Diffusion Large Language Models (DLLMs) have emerged as a compelling parallel decoding paradigm. However, the inherent full bidirectional attention at every denoising step incurs quadratic complexity, imposing severe latency bottlenecks in long-context scenarios. We investigate attention dynamics in DLLMs and propose Structured Sparse Attention (SSA), which uses a fast structure-wise algorithm to select and reuse geometric patterns for inference acceleration. SSA combines cached structured patterns with protection of the active decoding region to balance inference efficiency and output quality. Our investigation reveals a decrease in average adjacent-step attention changes later in denoising, together with vertical and diagonal geometric patterns that motivate sparse-pattern reuse; the trend varies across inputs, layers, and heads. Extensive experiments demonstrate that SSA delivers significant speedups while maintaining competitive quality against previous state-of-the-art methods on the LongBench and RULER benchmarks. At a 128K context length, SSA achieves an end-to-end inference speedup of approximately 2 × . These findings support structured-pattern reuse combined with decoding-region protection as an efficient, training-free approach to accelerating long-context DLLM inference.