Yuanzhong Xu

Gemini area lead (frameworks) 2024- Created the new-generation Gemini RL infra that enabled Gemini 2.5, Deep Think, Search AI Mode, Agents, etc. 2023-2024 Pretrain scaling. Defined the parallelism abstractions, and created novel techniques in improving performance.

Leading efforts on large model scaling at Google Deepmind. My work powers the training and serving of the most powerful large models at Google. Projects: Gemini: performance lead. Innovative large-scale parallelism. PaLM 2: infra lead for scaling, parallelism, and performance GSPMD: model partitioner shared by both JAX (pjit) and Tensorflow Several large-model frameworks, including paxml/praxis saxml: lightweight and efficient large model serving on JAX. Large model across domains: language, vision, audio, etc.

Google Brain Made GSPMD a key technology that powers many of Google's recent large models (Gemini, PaLM 1&2, GLaM, BigSSL, Parti, T5X, M4). Worked closely with researchers, and sometimes deep-dive into modeling itself (e.g., Parti)

Google Brain Giant models. GShard: https://arxiv.org/abs/2006.16668 XLA GSPMD partitioner: https://arxiv.org/abs/2105.04663 Weight update sharding: https://arxiv.org/abs/2004.13336 XLA and TPUs, TensorFlow and JAX

XLA compiler and TPUs: advanced low-level and high-level optimizations

Cross-datacenter traffic management. Cluster load testing framework.

Computer systems. Security.

Security infrastructure.

OS security. SGX evaluation.

Low-level messaging protocol.

Computer systems and security