Ying Zhang
Member Of Technical Staff at xAI
Palo Alto, California, United States
Summary
🤩
Rockstar🎓
Top SchoolYing Zhang is a GPU and accelerator-focused ML inference engineer with 10 years of experience building high-performance kernels, compilers, and distributed systems for production-scale models. Currently a Member of Technical Staff at xAI after senior roles at Meta where she co-authored FlashAttention-3 and led AITemplate improvements that outperformed TensorRT on key workloads, she blends low-level CUDA/kernel work with compiler and runtime optimizations. Her open-source contributions to PyTorch (Inductor, FP8, Triton/CUDA heuristics) and FBGEMM (embedding pruning and remapping kernels) demonstrate a track record of improving inference efficiency across frameworks and hardware backends. Earlier roles in large-scale SQL planners and streaming systems at Alibaba and Google give her deep systems and planner experience that informs performance tradeoffs in model serving. Colocated in Palo Alto, she is known for shipping pragmatic auto-tuning, fusion, and parallelism features that squeeze latency and cost out of real-world ML deployments.
10 years of coding experience
15 years of employment as a software developer
Shijiazhuang NO.2 High SchoolBachelor Computer Science, Bachelor Computer Science at Shanghai Jiao Tong University
English, Chinese
Github Skills (26)
pytorch10
performance-monitor10
performance-analytics10
machine-learning10
performance-measurement10
triton10
performance-analysis10
performance-tuning10
compiler10
cuda10
performance-monitoring10
induction10
optimization10
benchmark9
python9
Programming languages (6)
Github contributions (5)
pytorch/pytorch
Feb 2019 - Jul 2022
Tensors and Dynamic neural networks in Python with strong GPU acceleration
Role in this project:
ML Engineer
ML EngineerContributions:124 reviews, 9 commits, 40 PRs in 3 years 6 months
Contributions summary:Ying primarily contributes to the PyTorch framework, focusing on performance optimizations and the integration of new features. Their work includes implementing and refining functionalities related to the Inductor compiler, particularly for handling dynamic shapes and supporting advanced techniques like FP8 quantization and fusion optimizations. These changes involve modifications to Triton heuristics, CUDA kernels, and overall compilation strategies to improve performance and address specific issues in the PyTorch ecosystem. They also contribute to the CUTLASS backend and benchmark frameworks.
pythongpu-accelerationdeep-learninggpunumpy
pytorch/FBGEMM
Sep 2021 - Nov 2022
FB (Facebook) + GEMM (General Matrix-Matrix Multiplication) - https://code.fb.com/ml-applications/fbgemm/
Role in this project:ML Engineer
ML EngineerContributions:5 commits, 4 PRs in 1 year 2 months
Contributions summary:Ying focused on adding and improving embedding pruning support within the fbgemm library, specifically for use in inference. They implemented features related to pruning ratios, L2 norm-based pruning, and index remapping. The contributions included modifying the inference converter, adding pruning options to the benchmark tests, and implementing a new kernel for array-based index remapping. These changes aim to optimize performance and efficiency of embedding lookups.
matrix-multiplicationfacebookmultiplicationmatrixml-applications
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.