Leo (Linxiao) Li

I am a researcher focusing on large-scale LLM systems and AI infrastructure. Currently I work as an AI Infrastructure / Training Systems Researcher at AIGCode (蔻町科技).

My work focuses on large-scale LLM training systems. I study system-level optimizations to improve efficiency and hardware utilization, including Model FLOPs Utilization (MFU) improvement, communication–computation overlap, operator/kernel fusion, and runtime-level scheduling and memory optimization. I work closely with distributed training stacks to improve scalability, throughput, and cost-efficiency of foundation model training.

Previously, I was a Senior Researcher and core member at Huawei Cloud Architecture Innovation Lab (Cloud Lab), under CTO Fellow Jiongjiong Gu (Chief Architect), working on MoE inference, memory offloading, and cloud-scale AI infrastructure. I also collaborated with the Guangming Laboratory (Intelligent Recommendation & Resource Scheduling Group).

My long-term research goal is to design more efficient system architectures for large-scale AI models, improving scalability and resource efficiency of future AI infrastructure.

My research and engineering interests include:

• Efficient training and inference systems for large language models
• Model FLOPs Utilization (MFU) optimization and system efficiency
• Communication–computation overlap and operator / kernel fusion
• Distributed training systems and memory-efficient execution
• Scalable inference systems for large models (LLMs, MoE)

Check out my publications, CV, and gallery for more details.