Lei Wang
Software Engineer at Meta
Beijing, United States
Summary
🤩
Rockstar🎓
Top SchoolLei Wang is a software engineer with 8 years of experience specializing in compilers and performance optimization for ML workloads and accelerators. Currently on Meta's LLVM Server Compiler team, he focuses on AutoFDO/PGO and llvm-profgen to squeeze runtime efficiency from large systems. His open-source work on high-profile projects like Apache TVM and Microsoft NNFusion demonstrates deep hardware-aware compiler skills—adding TensorIR/CUDA/ROCm optimizations, AMD Matrix Core intrinsics, and FP16 codegen fixes for real-world models such as ResNet50. Previously at Huawei he engineered compiler frameworks for network processors, covering register allocation, instruction scheduling, and graph/branch transformations. He holds a PhD in computer science and balances research rigor with pragmatic engineering, often surfacing subtle correctness fixes and test coverage improvements that have outsized impact. Based in Beijing with a “create value” ethos, he blends low-level systems expertise with practical ML deployment experience.
8 years of coding experience
2 years of employment as a software developer
Master, Computer Science, Master, Computer Science at University of Chinese Academy of SciencesBachelor's degree, Computer Science, Bachelor's degree, Computer Science at Tianjin University
Github Skills (16)
cuda10
compiler10
machine-learning10
roc10
compiler-compiler10
c-language10
deep-learning10
deeplearning-ai10
cprogramming-language10
gpu10
compilation10
compile10
onnx9
tensorflow9
p89
Programming languages (16)
Github contributions (5)
microsoft/nnfusion
Apr 2022 - Nov 2022
A flexible and efficient deep neural network (DNN) compiler that generates high-performance executable from a DNN model description.
Role in this project:
ML Engineer
ML EngineerContributions:1 review, 43 commits, 19 PRs in 7 months
Contributions summary:Lei primarily contributed to improving and extending the functionality of the nnfusion compiler, focusing on adding new tests for different scatter operations in the TensorFlow frontend. They fixed a typo related to scattermin op, added codegen for fp16 in dot operations, and corrected the handling of float16 data within the ONNX import. Moreover, the user addressed issues in batch norm code generation and the import of the resnet50.onnx model with fp16.
tvmdeep-learningdeep-neural-networkcompilerneural-network
apache/tvm
Mar 2021 - Mar 2021
Open deep learning compiler stack for cpu, gpu and specialized accelerators
Role in this project:ML Engineer
ML EngineerContributions:23 reviews, 1 commit, 24 PRs in 1 day
Contributions summary:Lei primarily contributed to the development and optimization of the TVM compiler stack, specifically focusing on features related to TensorIR and CUDA/ROCm backends. They implemented support for L2 prefetch options and asynchronous copy operations in the TensorIR, improving performance in CUDA. The user also made significant changes to the ROCm target, including arch parsing and AMD Matrix Core support, which included the development of new tensor intrinsics for AMD GPUs, demonstrating deep understanding of hardware-specific optimization techniques.
metalvulkancompilertensoropencl
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.