Baizhou Zhang

Bachelor of Science, Intelligence Science and Technology

Beijing, China

Visiting Student

Berkeley, USA

Master of Science, Computer Science and Engineering

La Jolla, USA

Member of Technical Staff

Palo Alto, California, United States · On-site

• Member of Founding Team; Core developer and maintainer of SGLang.
• Co-led the optimization of GLM-5.2 model on Blackwell hardware. [Blog]
• Co-led the design and implementation of the parallelism strategy for Kimi-K3 Day 0 support (Prefill PP + Decode CP) [Blog]
• Maintainer of Prefill Context Parallelism feature for SGLang. Independently designed and implemented a cleaner version of PCP with CPStrategy abstraction. [Roadmap]
• Co-led the initial implementation of NVFP4 KV Cache in SGLang [Blog]
• DeepSeek V4 support and optimization
  • Implemented MTP speculative decoding and Trtllm MoE runner for DeepSeek V4 Day-0 support. [Blog]
  • Reviewed subsequent optimizations on DeepSeek V4 and helped increase its throughput to 5x compared with Day 0. [Blog]
• Implemented a tvm-ffi wrapper for SGLang's DeepGemm fork, and created a customized PyPI wheel sgl-deep-gemm for more flexible development and usage.

Open Source Developer (Among top 10 contributors)

California, United States · Remote · Part-time

• Core Developer and Maintainer of SGLang (among the top 10 committers).
• Led the initial support of deterministic inference in SGLang. [Blog] [X post] [量子位]
• Enhanced chunked prefill performance for the DeepSeek V3/R1 model by using a chunked prefix cache technique with MHA kernels, and achieved approximately 1.4x throughput improvement in long-context (32k) scenarios. [PR]
• Co-led the initial support of SGLang on Blackwell architecture, and deployment of Large-Scale Expert Parallelism on GB200 and GB300. [GB200 blog part 1] [GB200 blog part 2] [GB300 blog]
• Implementation of multiple attention backends for MLA, including FlashInfer backend ([PR1], [PR2]), and Flash Attention 3 backend ([PR]).
• Maintainer of LoRA feature. [Roadmap]
• Early development of srt-slurm, a framework that easily benchmarks models deployed with SGLang and NVIDIA Dynamo.

Software Engineer Intern

Santa Clara, California, United States · On-site

• CuDNN Heuristic Team. Supported heuristics for compiler-generated Blackwell GEMM kernels.

Software Engineer Intern

Beijing, China · On-site

• PaddlePaddle group. Developed the ML compiler CINN.

Software Engineer Intern

Beijing, China · On-site

• Developer of ColossalAI, a distributed training framework.

Qingke Community

Online

[Video]

RadixArk & Nvidia

San Jose, US

[Video] [X post] [Blog]

SGLang Shanghai Meetup / SGLang & Machine’s Heart

Shanghai, China

[WeChat Post]

Dynamo Day / Nvidia

Online

[Event] [Video]

Ant Open Source Meetup / SGLang & Ant Open Source

Hangzhou, China

[WeChat Post]

SGLang & Dynamo Meetup / SGLang & Nvidia

San Francisco, US

[Video]

Research Assistant

Beijing, China · On-site

• Applied QLoRA model quantization to cloud-device knowledge distillation.
• Supervised by Prof. Shanghang Zhang.

Research Assistant

Remote

• Researched music generation with VAE-based models.