Xingyuan Li
AI Framework Engineer at Intel Corporation
Shanghai, Shanghai, China
Summary
🤩
Rockstar🎓
Top SchoolXingyuan Li is an AI Framework Engineer at Intel with three years of hands-on experience optimizing ML runtimes and model tooling. Based in Shanghai and holding an MEng from Beijing University of Posts and Telecommunications, he focuses on integrating and improving compiler and conversion pipelines to accelerate LLM inference on Intel hardware. His open-source work includes substantive contributions to PyTorch’s Inductor compiler—adding tests and fixes for Intel GPU support and mixed-precision memory optimizations—and practical tooling in intel/ipex-llm that streamlines model conversion and CLI wrappers across major LLM families. He blends low-level performance engineering with usability improvements, reducing friction for deploying large models on diverse XPUs. Colleagues rely on him for pragmatic fixes that bridge research code and production-ready workflows.
3 years of coding experience
Master of Engineering - MEng, Master of Engineering - MEng at Beijing University of Posts and Telecommunications
Github Skills (15)
transformers10
machine-learning10
pytorch10
model-conversion10
deep-learning10
gpu10
python10
induction10
llm10
testing9
cli9
command-line9
command-line-interface9
autograd7
tensor7
Programming languages (7)
Github contributions (5)
intel/ipex-llm
Jul 2022 - Oct 2022
Accelerate local LLM inference and finetuning (LLaMA, Mistral, ChatGLM, Qwen, Mixtral, Gemma, Phi, MiniCPM, Qwen-VL, MiniCPM-V, etc.) on Intel XPU (e.g., local PC with iGPU and NPU, discrete GPU such as Arc, Flex and Max); seamlessly integrate with llama.cpp, Ollama, HuggingFace, LangChain, LlamaIndex, vLLM, GraphRAG, DeepSpeed, Axolotl, etc
Role in this project:
ML Engineer
ML EngineerContributions:48 reviews, 16 commits, 124 PRs in 3 months
Contributions summary:Xingyuan primarily focused on modifying and improving model conversion scripts for large language models within the Intel IPEX-LLM repository. They fixed bugs related to model conversion, specifically renaming model components for compatibility. The user also implemented command-line wrappers for various LLM families (llama, bloom, gptneox, chatglm, starcoder), enabling easier interaction with different models. Furthermore, they contributed to the setup process, including binary file management and dependency checks for improved usability.
llm-inferencellama2pythonfinetuningllama
pytorch/pytorch
Mar 2024 - Oct 2024
Tensors and Dynamic neural networks in Python with strong GPU acceleration
Role in this project:ML Engineer
ML EngineerContributions:22 reviews, 15 PRs, 64 comments in 7 months
Contributions summary:Xingyuan's commits focus on integrating and optimizing the Inductor compiler within the PyTorch ecosystem. Their contributions include reusing and generalizing existing Inductor test cases to support Intel GPUs. They've added new tests and fixed existing ones to ensure the Inductor compiler functions correctly for various optimization scenarios, particularly related to memory management and mixed-precision calculations. These changes involve modifying existing test files and adding new test cases to cover various operations and configurations within the PyTorch framework.
pythongpu-accelerationdeep-learninggpunumpy
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.