Nikolay Lyalyushkin
Staff Research Engineer at Cerebras
Munich, Bavaria, Germany
Summary
🤩
Rockstar🎓
Top SchoolNikolay Lyalyushkin is a Staff Research Engineer with 13 years of experience specializing in hardware-aware AI and inference optimization for Large Language Models and Generative AI. Based in Munich, he has driven quantization and compression advances at Intel—co-inventing a post-training LoRA correction and engineering INT8/INT4/FP4 workflows that materially boosted performance on Intel architectures and OpenVINO. He contributed core improvements to the widely used NNCF repository, refining Hessian-aware quantization initialization and mixed-precision handling to make compression more robust for production models. Comfortable bridging research and engineering, he has a track record of taking prototype algorithms into scalable toolchains (Torch FX, PyTorch, ONNX) and collaborating with hardware teams to adopt novel kernels. Notably, he reduced accuracy loss in quantization-aware training with absorbable LoRA adapters, halving previous degradation—an example of his focus on practical, measurable gains.
13 years of coding experience
4 years of employment as a software developer
Master's degree, Computational Machematics and Cybernetics, Master's degree, Computational Machematics and Cybernetics at State University of Nizhni Novgorod named after N.I. Lobachevsky (UNN)
Github Skills (6)
quantization10
pytorch10
machine-learning10
deep-learning10
tensorflow9
compression9
Programming languages (4)
Github contributions (5)
openvinotoolkit/nncf
Jun 2020 - Jan 2023
Neural Network Compression Framework for enhanced OpenVINO™ inference
Role in this project:
ML Engineer
ML EngineerContributions:1260 reviews, 109 commits, 314 PRs in 2 years 7 months
Contributions summary:Nikolay primarily contributed to improvements and refactoring within the HAWQ (Hessian Aware Quantization) precision initialization algorithm, a core component of the NNCF framework for neural network compression. Their work involved optimizing Hessian trace calculations, refactoring and debugging tools, and introducing features like compression ratio calculations. They also focused on handling frozen layers and addressing issues related to the application and integration of mixed-precision quantization techniques.
bertsemantic-segmentationmixed-precision-trainingtensorflowclassification
ljaljushkin/nncf_pytorch
Jun 2020 - Apr 2025
Contributions:19 reviews, 10 PRs, 979 pushes in 4 years 10 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.