Shouzheng Liu
AI GPU Performance Engineer at Modular
United States
Summary
🤩
Rockstar🎓
Top SchoolShouzheng Liu is an AI GPU Performance Engineer with six years of experience specializing in on-device LLM inference and GPU kernel optimization. Currently at Modular, he previously drove high-impact contributions to the widely used ggml-org/llama.cpp project, engineering Apple GPU matrix-multiplication kernels that reached 88% of theoretical peak and outperformed Apple's Metal Performance Shader. His work on concurrent dispatch and grouped-query attention reduced token-generation latency and sped quantized-model inference by up to 80%, demonstrating deep expertise in low-precision formats, parallel execution, and profiling-driven tuning. Trained as a physicist (PhD candidate at NYU, BS from USTC), he brings a researcher's rigor to production performance engineering and a knack for squeezing hardware efficiency out of constrained platforms.
6 years of coding experience
Doctor of Philosophy, Condensed Matter and Materials Physics, Doctor of Philosophy, Condensed Matter and Materials Physics at New York UniversityB.S., Physics, B.S., Physics at University of Science and Technology of China
Github Skills (10)
quantization10
c-language10
gm10
gml10
cprogramming-language10
gpu10
optimisation10
optimization10
linear-algebra9
machine-learning9
Programming languages (5)
Github contributions (5)
ggml-org/llama.cpp
Jul 2023 - Sep 2023
LLM inference in C/C++
Role in this project:
ML Engineer
ML EngineerContributions:16 reviews, 13 PRs, 57 comments in 1 month
Contributions summary:Shouzheng made significant contributions to optimizing the `llama.cpp` project for Metal (Apple GPU) hardware, focusing on performance enhancements and code size reduction. Their work involved implementing and refining matrix-vector and matrix-matrix multiplication kernels, along with modifications to the `rms_norm` kernel, leading to notable speed improvements, particularly for quantized models. Furthermore, the user introduced support for grouped-query attention (GQA) and implemented the concurrent dispatch of commands, enhancing the efficiency of model execution on Metal.
ggmlllama
lshzh-ww/llama.cpp
Jul 2023 - Aug 2023
Port of Facebook's LLaMA model in C/C++
Contributions:59 pushes, 16 branches in 1 month
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.