Bridging the Hardware Gap With Code Harnesses on the Hugging Face Kernels Hub - Ben Burtenshaw

This talk covers the development of code harnesses for standardizing kernel writing in the context of agentic coding. The speaker, Ben Burtenshaw from Hugging Face, presents an end-to-end experiment that benchmarks six harnesses across ten models using CUDA and Metal. Key topics include agent cost, kernel latency, VRAM usage, and end inference performance, highlighting how the Kernels Hub facilitates large-scale distribution. The talk also introduces two tools: the Kernels Hub, an infrastructure for creating reproducible kernels in the PyTorch ecosystem, and HF Skills, a library for defining and evaluating skills necessary for machine learning tasks related to kernel writing. This session is aimed at kernel writers and PyTorch developers seeking to enhance their understanding of high-performance kernel optimization.