DyLLM: Efficient Diffusion LLM Inference via Saliency-Based Token Selection and Partial Attention

arXiv 2603.08026·medium signal

Identifies salient tokens via attention context similarity and recomputes operations only for high-importance tokens while reusing cached activations elsewhere, achieving up to 9.6x higher throughput over baseline diffusion LLM inference while preserving accuracy. First practical throughput method for the emerging diffusion-based LLM class (Mercury, MDLM), which has previously lacked the inference optimization tooling available to autoregressive models.