Lizhuo Luo, Shenggui Li, Yonggang Wen, Tianwei Zhang
The paper introduces Dynamic Sliding Block (DSB), a training-free scheduling method for diffusion large language models that adapts to semantic difficulty, improving both quality and efficiency of text generation.
This research presents a new method called Dynamic Sliding Block (DSB) for improving text generation in diffusion large language models. These models are known for their ability to generate text efficiently by processing multiple parts of a sentence at once. However, the traditional way of scheduling these parts doesn't consider how difficult each part is, which can lead to mistakes. DSB adjusts the scheduling dynamically based on how challenging each part is, leading to better and faster text generation. The study shows that DSB, along with a caching mechanism called DSB Cache, enhances both the quality and speed of text generation.
