Reproduce “Denoising Surface: Modeling and Predicting Inference Cost for Diffusion LLM Serving” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haoyu Zheng, Fangcheng Fu, Binhang Yuan, Yongqiang Zhang, Liang Deng, Hao Wang, Yuanyuan Zhu, Xiao Yan and 1 more
DWS predicts computation for diffusion language-model requests before generation by estimating which output-block denoising steps will occur, then combining those probabilities with profiled execution costs. Its predictor runs on CPU. On test traces, it achieved 355.24 ms total-cost mean absolute error versus 517.04 ms for a total-steps predictor; online-chatbot gains were evaluated in a specific setup, and configuration changes may require retraining.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.