Research talk: Torchy: A tracing JIT compiler for PyTorch

Опубликовано 8 февраля 2022, 17:42

Speaker: Nuno Lopes, Principal Researcher, Microsoft Research Cambridge

This session will introduce current technologies to speed up training of machine learning models and their limitations, and it will also include a presentation of Torchy, a transparent solution to speed up PyTorch workloads through the means of JIT compilation. Machine learning models and their respective code keep getting more complex and dynamic. Traditional static compilation techniques (for example, TorchScript) can't handle these real-world, modern code bases. Furthermore, using techniques like TorchScript is risky and requires a very good understanding of the technology. The researchers present Torchy, a new tracing JIT compiler for PyTorch that works with any code base—improving its performance—and automates parallelism and distribution across multiple devices.

Learn more about the 2021 Microsoft Research Summit: Aka.ms/researchsummit