Yufeng Li
Principal Software Engineering Manager at Microsoft
Cupertino, California, United States
Summary
🤩
Rockstar🎓
Top SchoolYufeng Li is a Principal Software Engineering Manager based in Cupertino with eight years of experience leading high-performance ML inference work at Microsoft. He currently leads the ONNX Runtime GenAI initiative, driving CPU and GPU optimizations and delivering state-of-the-art transformer performance for models like Phi3, Llama, Gemma, Whisper and Stable Diffusion. His hands-on background spans efficient CUDA kernels, MLAS CPU library development (fp32/int8/int4) and quantized kernel work (QGemm/MatMulInteger), contributing directly to the widely used microsoft/onnxruntime project. Earlier roles in Bing and PowerPoint built his strengths in large-scale data pipelines and product-facing backend systems, blending research-grade performance engineering with product delivery. Known for squeezing maximal performance from hardware while keeping binaries lean, he partners across Bing, Office and Cognitive Services to put high-throughput GenAI into production.
9 years of coding experience
5 years of employment as a software developer
Master's degree Computer Engineering, Master's degree Computer Engineering at Beihang UniversityBachelor's degree Computer Science, Bachelor's degree Computer Science at Jilin University
Github Skills (15)
cuda10
matrix-multiplication10
quantization10
performance-tuning10
performance-monitor10
c-language10
performance-analysis10
deep-learning10
performance-monitoring10
performance-analytics10
cprogramming-language10
linear-algebra10
performance-measurement10
transformers9
openmp5
Programming languages (6)
Github contributions (5)
microsoft/onnxruntime
Dec 2018 - Jan 2023
ONNX Runtime: cross-platform, high performance ML inferencing and training accelerator
Role in this project:
Back-end & Performance Engineer
Back-end & Performance EngineerContributions:2 releases, 1690 reviews, 492 commits in 4 years 2 months
Contributions summary:Yufeng primarily focused on optimizing the performance of ONNX Runtime's deep learning inference capabilities. Their contributions included implementing and optimizing kernels for quantized matrix multiplication (QGemm, MatMulInteger) to leverage hardware acceleration, enabling the use of prepacked buffers for better memory usage, and improving the accuracy and efficiency of the attention mechanism, particularly for models using sequence processing, such as those used in Transformers. They also addressed bugs related to edge cases.
machine-learningonnxruntimedeep-learningonnxneural-networks
yufenglee/flash-attention
Apr 2023 - Nov 2025
Fast and memory-efficient exact attention
Contributions:1 PR, 2 pushes in 2 years 7 months
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.