Aaron Shi
Software Engineer, IG ML at Meta
Old Toronto, Ontario, Canada
Summary
🤩
Rockstar🎓
Top SchoolAaron Shi is a software engineer with a decade of experience specializing in compiler toolchains, GPU performance, and ML infrastructure, currently improving recommendation model efficiency on Instagram's ML team at Meta. He has deep backend and systems expertise from leading PyTorch profiling (Kineto) and memory tooling (Memento) work to contributing substantive HIP/ROCm and Clang backend improvements at AMD. His open-source contributions show hands-on mastery of build systems, CI/CD, GPU compatibility and low-level performance fixes—skills that bridge compiler internals and production ML workloads. Based in Toronto with an MEng from the University of Toronto, he blends research-grade technical depth with practical engineering that scales across large infra and hardware-software integration.
10 years of coding experience
5 years of employment as a software developer
Master of Engineering - MEng Computer Engineering, Master of Engineering - MEng Computer Engineering at University of Toronto
English
Github Skills (40)
pytorch10
c-language10
performance-analytics10
performance-monitor10
gpgpu10
system-programming10
python10
code-profiling10
llvm10
low-level10
cmake10
cicd10
hip10
roc10
low-level-programming10
Programming languages (10)
Github contributions (5)
pytorch/kineto
Sep 2021 - Jan 2023
A CPU+GPU Profiling library that provides access to timeline traces and hardware performance counters.
Role in this project:
Back-end Developer & Performance Engineer
Back-end Developer & Performance EngineerContributions:206 reviews, 71 commits, 149 PRs in 1 year 4 months
Contributions summary:Aaron primarily focused on improving the Kineto library, which is a CPU+GPU profiling tool for PyTorch. Their contributions involved enhancing the logging capabilities of the library, adding support for collecting and saving logs to trace files and Scuba, and refactoring the logging infrastructure. They also addressed performance issues by optimizing activity buffer sizes and improving the efficiency of the codebase to reduce overhead. The user's work included modifying C++ code and related headers.
cudacputracesprofilingtimeline
ROCm/hip
Sep 2017 - Jul 2021
HIP: C++ Heterogeneous-Compute Interface for Portability
Role in this project:Back-end Developer & System Architect
Back-end Developer & System ArchitectContributions:157 commits, 66 PRs, 52 pushes in 3 years 10 months
Contributions summary:Aaron primarily focused on improving the HIP (Heterogeneous-Compute Interface for Portability) library by fixing bugs and improving compatibility for various target architectures. Their work involved correcting math function implementations and adjusting target naming conventions related to AMDGPU. Furthermore, the user refactored internal API calls and refactored memory fence functions for memory scope and memory order. The user's commits demonstrate a deep understanding of the underlying architecture and the porting aspects of the code.
cudaheterogeneousgpuportabilityhip-runtime
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.