Hi team,
I am exploring how to integrate a custom backend with PyTorch. After reading about custom backends and The Custom Operators Manual, it looks like modern custom backend support often relies on computational graphs. However, in scenarios where there are more dynamic operations and a computational graph cannot be extracted, it seems that integrating a custom backend in eager mode is necessary.
Is my understanding correct? If so, could you kindly share the latest recommended way to integrate a custom backend in eager mode with PyTorch? Thank you.