Jinghan Yao
English name: Murphy ('s law)

Phone: 425-543-4197
Email: yao.877@osu.edu
Hi, I am currently in my second year of Ph.D. studies at The Ohio State University, advised by Dr. Dhabaleswar K. (DK) Panda, focusing the research on developing advanced high-performance computing (HPC) and networking strategies specifically for artificial intelligence (AI) models and applications. My work is particularly geared towards large-scale parallel inference on AI models, with a special focus on foundation models in Natural Language Processing (NLP) and computer vision. This includes the design of efficient GPU kernels and the optimization of communication strategies for distributed inference systems.
Before joining OSU, I was a research assistant at Fudan University, advised by Prof. Li Zhang at Zhang Vision Group . My research focus was on designing efficient Transformers for computer vision and NLP tasks.
news
Feb 12, 2024 | My work “Flover”, will be presented on NVIDIA GTC 2024 keynote session. Look forward to seeing you at San Jose in March. |
---|---|
Dec 23, 2023 | My paper “Exploiting Inter-Layer Expert Affinity for Accelerating Mixture-of-Experts Model Inference” has been accepted by IPDPS’24. |
Nov 11, 2023 | Join the amazing SC'23 conference as student volunteer, at Denver, Colorado. Together with our extraordinary team at OSU! |
Oct 04, 2023 | My paper “Flover: A Temporal Fusion Framework for Efficient Autoregressive Model Parallel Inference” has been accepted to HiPC’23. |
selected publications
- Exploiting Inter-Layer Expert Affinity for Accelerating Mixture-of-Experts Model InferenceAdvances in 38th IEEE International Parallel & Distributed Processing Symposium (IPDPS 24), 2024
- Flover: A Temporal Fusion Framework for Efficient Autoregressive Model Parallel InferenceAdvances in 30th IEEE INTERNATIONAL CONFERENCE ON HIGH PERFORMANCE COMPUTING, DATA, & ANALYTICS (HiPC 23), 2023
- Soft: Softmax-free transformer with linear complexityAdvances in Neural Information Processing Systems (NeurIPS 21), 2021
- SPRNet: single-pixel reconstruction for one-stage instance segmentationIEEE transactions on cybernetics, 2020