Kacent Huang

Google Summer of Code 2026 Contributor

email: fogsong233[at]gmail[dot]com

Education: Computer Science, Nanjing University, Nanjing, China

Ongoing project: Clad as a First-Class Gradient Engine in LibTorch
This proposal targets the HSF 2026 idea Clad as a first-class gradient engine in LibTorch, a 350-hour project intended to make compiler-generated gradients available from LibTorch and, eventually, easier to reuse from the ROOT ecosystem. My goal is to build a practical and well-scoped integration path that allows a LibTorch C++ training loop to call Clad-generated derivative code for a small but meaningful subset of workloads. The primary implementation path is to wrap Clad-generated backward logic inside torch::autograd::Function, because the PyTorch C++ frontend explicitly supports custom autograd functions and presents them as the standard way to integrate optimized forward and backward code in extensions. If time permits and the API proves cleaner, I will also evaluate a custom-operator path based on the PyTorch C++ extension mechanisms. This proposal is intentionally scoped as a proof of concept, not as a general replacement for LibTorch autograd. A realistic first milestone is to differentiate selected training kernels or small model components written in C++, expose them to LibTorch, and measure where compiler-generated derivatives are correct, maintainable, and competitive. This produces a concrete result for mentors and also establishes a clear baseline for future work on broader operator coverage, deeper ROOT integration, or GPU support.

Project Proposal: URL

Mentors: Aaron Jomy, David Lange, Vassil Vassilev

Presentations

---

---