April 28, 2026 · papers
dLLM-Serve accepted at ICS 2026
Diffusion language models generate text by iterative refinement, and the memory traffic that requires has kept them out of production serving. dLLM-Serve closes that gap. The paper appears at the 40th ACM International Conference on Supercomputing in Belfast. Congratulations to Jiakun Fan, Yiyuan Zhang and Xiangchen Li.