Sparse-dLLM: Accelerating Diffusion LLMs with Dynamic Cache Eviction

TL;DR

Sparse-dLLM introduces a training-free, dynamic cache eviction method that leverages token saliency in diffusion LLMs to significantly improve inference throughput while maintaining performance and memory efficiency.

Contribution

It presents the first training-free framework combining dynamic cache eviction with sparse attention for diffusion LLMs, enhancing inference efficiency.

Findings

Achieves up to 10× higher throughput than vanilla dLLMs.

Maintains comparable performance and peak memory costs.

Outperforms previous methods in efficiency and effectiveness.

Abstract

Diffusion Large Language Models (dLLMs) enable breakthroughs in reasoning and parallel decoding but suffer from prohibitive quadratic computational complexity and memory overhead during inference. Current caching techniques accelerate decoding by storing full-layer states, yet impose substantial memory usage that limit long-context applications. Our analysis of attention patterns in dLLMs reveals persistent cross-layer sparsity, with pivotal tokens remaining salient across decoding steps and low-relevance tokens staying unimportant, motivating selective cache eviction. We propose Sparse-dLLM, the first training-free framework integrating dynamic cache eviction with sparse attention via delayed bidirectional sparse caching. By leveraging the stability of token saliency over steps, it retains critical tokens and dynamically evicts unimportant prefix/suffix entries using an attention-guided strategy. Extensive experiments on LLaDA and Dream series demonstrate Sparse-dLLM achieves up to 10$\times$ higher throughput than vanilla dLLMs, with comparable performance and similar peak memory costs, outperforming previous methods in efficiency and effectiveness. The code is available at https://github.com/OpenMOSS/Sparse-dLLM.