Juncheng Gu

About Me

I am working on Machine Learning Systems at SystemsResearch@Google. Prior to that, I worked at ByteDance as a research scientist. I obtained my Ph.D. degree in Computer Science and Engineering from University of Michigan, Ann Arbor, co-advised by Prof. Kang G. Shin, and Prof. Mosharaf Chowdhury.

I am interested in building efficient software systems for LLMs. My current work focuses on LLM serving — analyzing and optimizing internal serving solutions, and building disaggregated serving platforms that leverage both TPUs and GPUs for cloud customers. I am also building a disaggregated reinforcement learning training system that runs on elastic and heterogeneous resources.

At ByteDance, I contributed to large-scale training systems for both traditional ML and emerging LLMs, notably developing internal PyTorch distributions with customized optimizations. During my Ph.D. study, I focused on resource management for deep learning clusters relying on low-latency RDMA networks (thesis), and built the first memory disaggregation system over RDMA networks.

Publications