Yuyue Li 李宇越
Undergraduate Student (2024 intake)
National Elite Institute of Engineering, Chongqing University
重庆大学国家卓越工程师学院 · Chongqing, China
Research interests
- Efficient inference — KV cache design and long-context attention: what a model actually needs to attend to when the context gets long.
- Generative modeling — image generation and autoregressive generative models.
- LLM agents & tooling — agent memory, long-horizon task execution, and the harnesses that make them work in practice.
- Efficient architectures — mixture-of-experts offloading and GPU–CPU hybrid execution: the engineering side of making very large models run on hardware people actually have.
Most of this work lives in private repositories and in projects not yet published, so the public repositories below are only a slice of it.
Projects
WarpMoE
A GPU–CPU hybrid offloading engine that runs DeepSeek-scale MoE models far beyond available VRAM — 83–119 tok/s on a single A100, across 24–80 GB VRAM tiers.
Contact
- Email — 20242803@stu.cqu.edu.cn
- GitHub — @Mr-Skeleton-Max