xmu-xiaoma666/External-Attention-pytorch ? reverse-engineered prompt

Build me a PyTorch toolkit for visual deep learning that collects a bunch of reusable blocks in one place, especially attention modules, backbone models, MLP blocks, re parameterized blocks, and a few convolution variants.

I want it to be easy to install as a package, and also easy to use directly from the repo. Please include simple example code that shows how to import one of the modules, pass in a tensor, and see the output shape. A small demo script is enough, but it should be clear enough that someone new to the ideas can try them without digging through paper implementations.

Make the code organized and readable, with each module split cleanly so it’s easy to find the core logic. If there are any small compatibility issues with current PyTorch docs or import paths, please fix those too, and keep the README usage examples working from both the installed package and the cloned repo.