A team consisting of CSE professor Zizhong Chen and PhD student Yujia Zhai, together with researchers from ByteDance and NVIDIA received best paper award at IPDPS 2023, one of the top conferences on parallel and distributed processing.
The paper is titled "ByteTransformer: A High-Performance Transformer Boosted for Variable-Length Inputs".
Transformers have become keystone models in a variety of deep learning applications, such as natural language processing, neural machine translation, and recommendation systems. This paper presents
ByteTransformer, a high performance GPU-accelerated transformer optimized for variable-length inputs. The research is currently being used to power all of ByteDance's in-house search businesses, including TikTok, Douyin, Xigua Video, Toutiao Search Portal, and Magellan (an export-oriented e-commerce platform), serving over a billion active users worldwide every day.