dmlc/dgl ? reverse-engineered prompt

Build me a Python library for deep learning on graphs that feels easy to use for people who want to train graph neural networks without dealing with all the low level details.

It should let me create graphs, attach features to nodes and edges, and run message passing style models on CPU or GPU. I want it to work with the major deep learning frameworks, so the rest of the app can stay in whatever stack the user already has. Please include a simple command line way to train and try common GNN models, plus a set of clear examples and tutorials that show how to get started fast.

Also make it ready for large graphs, with support for distributed and multi GPU training if possible. I’d like the project to feel polished, with good docs, runnable examples, and a few benchmark style scripts so people can compare models and performance. If you need to look up current docs online while building it, go ahead.