主讲人
Xuefei Ning Tsinghua University 时间 2026年6月4日 星期四 下午 14:00-15:00 地点 学院104会议室
Abstract Large-scale generative models have demonstrated strong capabilities in language and visual generation, but their inference often incurs substantial computational and memory costs. Due to latency, privacy, deployment constraints, and cost considerations, efficiently deploying such models on resource-constrained devices remains a key challenge for AIGC applications. Biography Xuefei Ning is an Assistant Researcher at Tsinghua University. Xuefei received the B.Eng. and Ph.D. degrees from the Department of Electronic Engineering, Tsinghua University, in 2016 and 2021, respectively. Xuefei’s past research interests have primarily focused on efficient deep learning, including model compression, neural architecture design, and efficient sampling algorithms for generative models. Xuefei’s work has received over 4,300 citations on Google Scholar. Xuefei has served as a Senior Area Chair for ACL and EMNLP, and as an Area Chair for CVPR, ICLR, NeurIPS, and COLM. Xuefei is the author of the book Efficient Deep Learning: Model Compression and Design. The website of the research team Xuefei led is: https://nics-effalg.com/
In this talk, I will present a line of my research on improving the efficiency of generative model inference at both the algorithm and model levels. At the algorithm level, I develop faster sampling strategies to mitigate the inherent inefficiencies of different generative paradigms, including parallel generation for autoregressive models and few-step sampling schedules for diffusion models. At the model level, I investigate a range of compression and inference optimization techniques, including quantization, weight pruning, attention sparsification, and activation sharing. These methods aim to reduce computational and storage overhead while preserving generation quality and task performance. Finally, I will briefly discuss my ongoing research interests.
