Liang Depeng
研发工程师 at 北京一流科技有限公司
Guangzhou City, Guangdong Province, China
Summary
🤩
Rockstar🎓
Top SchoolLiang Depeng is a seasoned R&D engineer with 12 years of experience specializing in deep learning frameworks, model deployment and performance optimization, currently based in Guangzhou. He has driven OneFlow framework adaptations across multiple chips (ROCm, Cambricon, and Sugon/Suiyuan), single-handedly delivering ROCm support to run BERT/GPT multi-node training and leading operator porting and quantized-model verification on specialized hardware. Previously he optimized mobile inference for Bigo Live—covering ARM CPU operator acceleration, model quantization, and real-time segmentation and style-transfer deployments—and contributed bug fixes and examples to Apache MXNet. An active open-source contributor, his OneFlow work spans QAT, einsum, normalization ops and vision-model compatibility tests, reflecting deep familiarity with PyTorch internals and CUDA optimization. He combines systems-level engineering with practical ML model engineering, and maintains public code and technical posts on GitHub and Zhihu.
12 years of coding experience
Master's degree, 计算机科学与技术, Master's degree, 计算机科学与技术 at Sun Yat-Sen University
Chinese, 广州话, Chinese
Github Skills (18)
pytorch10
python10
evaluation10
machine-learning10
ml10
scala10
deep-learning10
deep-neural-networks10
neural-network10
quantization10
metric10
c-language9
cprogramming-language9
cuda9
computer-vision8
Programming languages (5)
Github contributions (5)
Oneflow-Inc/oneflow
Jun 2020 - Jan 2023
OneFlow is a deep learning framework designed to be user-friendly, scalable and efficient.
Role in this project:
ML Engineer
ML EngineerContributions:718 reviews, 345 commits, 127 PRs in 2 years 6 months
Contributions summary:Liang contributed code related to a deep learning framework designed for user-friendliness, scalability, and efficiency, indicated by the "OneFlow" repository. The commits focused on implementing and testing core features of the framework, specifically around quantization-aware training (QAT), and implementing operations such as einsum, flatten, instance normalization and batch normalization. The user also worked on improving existing operations like `roll` and `matmul`, and added more vision model compatibility tests.
caffe2tensorflowframework-learningchainerdeep-learning-framework
apache/mxnet
Mar 2016 - Mar 2018
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:Back-end Developer
Back-end DeveloperContributions:56 commits, 70 PRs, 221 comments in 1 year 11 months
Contributions summary:Liang primarily contributed to fixing implementation bugs related to the "mae" and "mse" evaluation metrics within the Scala package. These fixes involved modifications to the `EvalMetric.scala` file, indicating a focus on improving the accuracy and correctness of the deep learning model evaluation. Further contributions included integrating the neural-style example and enhancing the Scala examples by implementing run scripts. This suggests a role in refining the library's core functionality and usability for users.
pythonschedulerdataflowmutationdata-science
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.