Lightning Talk: Coding Agents for Compiler Construction: Beyond the... Reza Rahimi & Stefan Krassin

This talk focuses on coding agents as engineering tools for compiler construction, specifically within the context of machine learning compilers that utilize PyTorch. The speakers, Reza Rahimi and Stefan Krassin, address the complexities of optimizing GPU kernels due to evolving hardware architectures and the increasing demand for efficient kernel generation. They present a novel compiler architecture designed as a PyTorch add-on, which processes PyTorch models or FX graphs to produce executable artifacts. The session highlights how these coding agents can analyze model mathematics and hardware specifications to create optimized kernels tailored to specific targets, bridging the gap between model developers and hardware capabilities.