Ligeng Zhu
Researcher at NVIDIA
Cambridge, Massachusetts, United States
Summary
🤩
Rockstar🎓
Top SchoolLigeng Zhu is a machine learning and computer vision researcher with 11 years of experience, currently working at NVIDIA on efficient large models after research positions at MIT, SFU, and ZJU. His work bridges research and engineering—contributing to influential open-source projects like ICLR-winning Once-for-All and the TVM compiler—focusing on model efficiency, deployment, and reliable tooling. He has a strong practical bent for production readiness, demonstrated by packaging projects for PyPI, improving FLOP/MAC counters, and optimizing data pipelines via LMDB tooling. Holding a PhD-level trajectory at MIT and roots in computer science from SFU and Zhejiang University, Ligeng combines deep academic rigor with hands-on systems improvements that make large models practical at scale. An under-the-radar strength is his attention to correctness in low-level compiler and operator shape checks, which reduces subtle runtime failures in deployment.
11 years of coding experience
4 years of employment as a software developer
Bachelor’s Degree, Computer Science, Bachelor’s Degree, Computer Science at Zhejiang UniversityDoctor of Philosophy - PhD, EECS, Doctor of Philosophy - PhD, EECS at Massachusetts Institute of TechnologyBachelor’s Degree, Computer Science, Bachelor’s Degree, Computer Science at Simon Fraser University
English, Chinese, Japanese
Github Skills (38)
continuous-deployment10
tvm10
pytorch10
c-language10
preprocessing10
python10
preprocess10
lmdb10
relay-modern10
machine-learning10
ml-deployment10
data-preprocessing10
data-loading10
deeplearning-ai10
compiler-compiler10
Programming languages (13)
Github contributions (5)
Lyken17/Efficient-PyTorch
Mar 2018 - Jan 2020
My best practice of training large dataset using PyTorch.
Role in this project:
ML Engineer
ML EngineerContributions:21 commits, 3 PRs, 18 pushes in 1 year 9 months
Contributions summary:Ligeng contributed to the project by creating scripts to convert image folders into LMDB datasets, crucial for efficient data loading in PyTorch. They also incorporated ImageNet training code from PyTorch tutorials, indicating an attempt to adapt or integrate existing training methodologies within the repository's framework. Further contributions include the addition of tensorpack scripts for LMDB creation and argparse for improved command-line argument parsing. This suggests the user is focused on optimizing data handling and training processes for the repository's focus on efficient PyTorch training.
pytorchbest-practicedeep-learningdatasetmachine-learning
Lyken17/pytorch-OpCounter
Apr 2018 - Sep 2022
Count the MACs / FLOPs of your PyTorch model.
Role in this project:ML Engineer
ML EngineerContributions:5 reviews, 100 commits, 25 PRs in 4 years 6 months
Contributions summary:Ligeng contributed to the `pytorch-opcounter` repository, which focuses on counting FLOPs/MACs of PyTorch models. Their contributions primarily involved enhancing the counting mechanisms for various PyTorch operators, including the addition of zero-op placeholders for operators like MaxPool and Dropout. They also fixed bugs in the counting formula and added support for upsample calculations. Furthermore, the user added an evaluation script and updated it to measure the FLOPs and Params of various famous models.
pytorchmacsflopspytorch-model
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.