hkproj/pytorch-transformer ? reverse-engineered prompt

Build me a PyTorch project that implements the Transformer from Attention Is All You Need, with the whole thing easy to run in notebooks.

I want a working training flow, data loading, model code, and translation inference, plus a couple of notebooks that show how to train locally, train in Colab, test inference, and try beam search. It should also include an attention visualization notebook so I can see what the model is looking at.

Make it feel like a complete small demo project, not just the model class. I want clear scripts for training and translating, and sensible config settings I can tweak without digging through a lot of code. If anything is unclear, use current docs online and keep the setup simple and beginner friendly.