Hao Zhang

Hao Zhang is dedicated to tackling challenges such as the high training and inference costs of large language models (LLMs) and the closed nature of technical ecosystems. He has made a series of groundbreaking contributions in distributed machine learning systems, LLM optimization and deployment, and open ecosystem construction.

During his doctoral research, Hao Zhang proposed the core concept of “machine learning parallelism being adaptive, composable, and automatic” and built the world's first GPU parameter server, providing revolutionary support for large-scale deep learning. His subsequent work on distributed scheduling, Pollux, broke through the limitations of traditional static methods with a dynamic scheduling mechanism, significantly improving the efficiency of training large-scale deep learning models. This work earned him the Jay Lepreau Best Paper Award at the top-tier systems conference OSDI 2021.

Entering the era of Large Language Models, Hao Zhang expanded his research focus to LLM training and inference optimization. He played a core role in developing several key technologies such as Alpa (scalable LLM training), vLLM, and DistServe (efficient LLM serving). These innovations leverage novel parallelism strategies, memory management techniques, and dynamic resource scheduling to significantly reduce inference costs and increase service throughput, thus facilitating the practical deployment of large models.

Furthermore, as a co-founder of LMSYS.org, Hao Zhang actively promotes an open AI ecosystem. He led the development of Chatbot Arena and LLM-as-a-Judge, among the world's most influential open LLM evaluation platforms. These platforms foster the prosperity of the open-source AI ecosystem through automated, transparent evaluation. In the future, he aims to advance the multi-task adaptability of AI systems, optimize the architecture of large-scale intelligent systems, and continue expanding the open LLM ecosystem to make the large model ecosystem fully public and universally accessible.