Wynton Wu

@DeepSeek-AI·@Peking University

ytwu776@foxmail.com

👋 Hi — glad you stopped by.

  • 🛠️ Building LLM infrastructure at DeepSeek-AI — our in-house RL training and inference frameworks, think verl and vLLM.
  • 🎓 Started my Ph.D. at Peking University in 2025, advised by Prof. Xin Jin. Earned my B.S. there the same year, working on cloud computing and cloud-native infrastructure.
  • 📬 Drop me an email anytime — I'm always curious about what people are building from scratch, and happy to get involved.

Experiences

DeepSeek-AI

DeepSeek-AI
2025/08-Present

I build and maintain the RL training and inference infrastructure behind the next generation of DeepSeek models, delivering fast and reliable systems to our post-training researchers. The work spans the full stack — from low-level performance optimization to the application logic on top.

Tencent WXG

Tencent WXG
2025/02-2025/07 @Beijing
Research Intern

I work with the most excellent peers and engineers to build incredible inference system for SOTA large language models like DeepSeek-R1. Our system deployed for WeChat serves a billion of users.

We also work closely with open-source communities, such as SGLang and vLLM, to contribute our industrial insight and optimizations for extreme large deployment.

Syslab, University of Washington

Syslab, University of Washington
2024/06-2024/10 @Seattle
Research Intern

Microsoft Research Asia

Microsoft Research Asia
2023/04-2023/07 @Beijing
Research Intern

Teaching

Talks

Publications

  1. DualPath: Accelerating Agentic LLM Inference by Harvesting Disaggregated KV-Cache Storage I/O

    Yongtong Wu, Shaoyuan Chen, Rilin Huang, Yixuan Tan, Yinmin Zhong, Mingxing Zhang, Xin Jin, Panpan Huang

    SIGCOMM 2026, accept rate: 85/445=19.1%

    [arxiv] [SIGCOMM]

  2. DeepSeek-V3.2 / V4 / V4.1 Technical Reports

    DeepSeek-AI

    arXiv

    [V3.2] [V4] [V4.1-Flash]

  3. High-level Programming for Application Networks

    Xiangfeng Zhu, Yuyao Wang, Banruo Liu, Yongtong Wu, Nikola Bojanic, Jingrong Chen, Gilbert Bernstein, Arvind Krishnamurthy, Sam Kumar, Ratul Mahajan, Danyang Zhuo

    NSDI 2025, accept rate: 55/401=13.7%

  4. RB^2: Narrow the Gap between RDMA Abstraction and Performance via a Middle Layer

    Haifeng Sun, Yixuan Tan, Yongtong Wu, Jiaqi Zhu, Qun Huang, Xin Yao, Gong Zhang

    INFOCOM 2024, accept rate: ~20%

    [PDF]