Jun Su
Principal Development Manager at 微软
Shanghai, Shanghai, China
Summary
🤩
Rockstar🎓
Top SchoolJun Su is a Principal Development Manager at Microsoft with 15 years of engineering and leadership experience building highly scalable web and distributed systems, currently leading the effort to bring MySQL to Azure. He combines deep hands-on expertise from kernel and CPU-level optimization to cloud services and web-scale architecture, having started and grown cross-border engineering teams in Shanghai that collaborate with US and global sites. Jun has a proven track record shipping large-scale services at Microsoft since 2001, including founding the HPC China team and driving production Azure services to GA. An active open-source contributor, his work spans ML frameworks (MXNet, ggml/llama.cpp), low-level systems (FreeBSD kernel, Dolphin emulator) and embedded SDR projects, reflecting rare full-stack fluency. He’s especially skilled at bridging platform differences—Windows/*nix, CPU/GPU—and translating deep technical trade-offs into product strategy and engineering practices. Colleagues rely on him for tackling the hardest technical challenges while scaling teams and processes across time zones.
15 years of coding experience
12 years of employment as a software developer
Computer Science, Computer Science at Fudan University
English, Chinese
Github Skills (60)
pytorch10
opencl10
code-optimization10
c-language10
python10
back-end-development10
elixir10
sdl10
memory-management10
mxnet10
machine-learning10
build-system10
cross-platform10
cmake10
batchfile10
Programming languages (19)
Github contributions (5)
ggml-org/llama.cpp
Mar 2023 - Oct 2024
LLM inference in C/C++
Role in this project:
Backend Developer
Backend DeveloperContributions:60 reviews, 44 PRs, 9 pushes in 1 year 7 months
Contributions summary:Jun primarily focused on optimizing and enhancing the `ggml` library, which is used for LLM inference. Their contributions included optimizing the `rope` function within `ggml`, as well as improving OpenCL support by checking for FP16 support and optimizing memory allocation. Additionally, they refactored code to leverage mmap for offloading tensors and fixed several build and runtime issues within the CUDA and general ggml code. They also removed support for shards weight files.
cppllm-inferenceggml
apache/mxnet
Dec 2016 - Feb 2017
Lightweight, Portable, Flexible Distributed/Mobile Deep Learning with Dynamic, Mutation-aware Dataflow Dep Scheduler; for Python, R, Julia, Scala, Go, Javascript and more
Role in this project:ML Engineer
ML EngineerContributions:31 commits, 41 PRs, 125 comments in 2 months
Contributions summary:Jun made several contributions focused on improving the MXNet deep learning framework. Their work included fixing bugs related to the RCNN monitor and segment fault in the profiler. They also improved the build system, reduced the default learning rate for MNIST, and converted code to the new programming API. Further enhancements included adding features for the image classification examples and improvements in profiling.
dataflowdeep-learninggojavascriptjulia
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.