Xinya Zhang
Member Of Technical Staff at The University of Texas at Austin
Austin, Texas, United States
Summary
👤
Senior🎓
Top SchoolXinya Zhang is a Member of Technical Staff at AMD and a PhD-trained computer scientist with 13 years of experience specializing in motion planning with deep learning and hardware-accelerated ML. Based in Austin, she combines research and teaching roles at UT Austin with hands-on engineering work, contributing performance-critical features to flagship open-source projects like PyTorch and ONNX Runtime. Her contributions focus on enabling and optimizing attention mechanisms and ROCm execution for AMD GPUs—work that demands deep kernel-level knowledge and cross-platform adaptation. Xinya’s background spans academia and industry, pairing rigorous doctoral research with pragmatic system optimizations that measurably improve ML inference and training on emerging AMD architectures.
13 years of coding experience
7 years of employment as a software developer
Doctoral, Computer Science, 3.867, Doctoral, Computer Science, 3.867 at The University of Texas at AustinMaster's degree, Computer Science, 3.45, Master's degree, Computer Science, 3.45 at Fudan University
Chinese, English
Github Skills (22)
pytorch10
machine-learning10
onnx10
hardware-acceleration10
flash10
roc10
deep-learning10
self-attention10
gpu10
multihead-attention10
c-language9
python9
optimisation9
tensor9
optimization9
Programming languages (7)
Github contributions (5)
pytorch/pytorch
Jul 2022 - Jul 2022
Tensors and Dynamic neural networks in Python with strong GPU acceleration
Role in this project:
ML Engineer & Back-end Developer
ML Engineer & Back-end DeveloperContributions:55 reviews, 2 commits, 37 PRs in 1 day
Contributions summary:Xinya's primary contributions involve enabling and optimizing Flash Attention and other SDPA (Scaled Dot-Product Attention) features within the PyTorch framework, specifically targeting the ROCm platform. They've focused on integrating AOTriton kernels, addressing performance regressions, and adding support for various architectures like MI200 and MI300X. This includes implementing the necessary backend adaptations, ensuring compatibility, and addressing limitations to enhance the performance of attention mechanisms.
pythongpu-accelerationdeep-learninggpunumpy
microsoft/onnxruntime
Jul 2022 - Aug 2022
ONNX Runtime: cross-platform, high performance ML inferencing and training accelerator
Role in this project:ML Engineer
ML EngineerContributions:12 reviews, 6 commits, 9 PRs in 24 days
Contributions summary:Xinya primarily contributed to the ROCm (AMD) execution provider for ONNX Runtime. Their work involved enabling and optimizing various ONNX operators, including GridSample, MatMulInteger, InstanceNormalization, BatchNormalization, LRN, AveragePool, GlobalAveragePool, MaxPool, GlobalMaxPool, NGramRepeatBlock, LongformerAttention, and DecoderAttention, specifically targeting AMD GPUs. Furthermore, the user made changes to adapt existing code and leverage the ROCm/MIOpen fusion API, enhancing performance and supporting new features. The contributions demonstrate expertise in hardware acceleration and machine learning model optimization on AMD hardware.
runtimetrainingtensorflowai-frameworkaccelerator
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.