PyTorch has become the de facto standard for development and research in deep learning. Among the many factors contributing to its popularity is the wide array of customization hooks it provides. These extension mechanisms allow developers to build new functionality on top of PyTorch while maintaining compatibility with its core backend features—a powerful capability for engineers, researchers, and curious hackers, both in-core and downstream. In this talk, we’ll explore various ways to extend PyTorch and present concrete examples of these techniques in action.