[smcin]

> For clarity, this is ONLY the forward pass of the model. There's no training code, batching, kv cache for efficiency, GPU support, etc ...

Neat, but please add one-line comments/docstrings where these missing bits would go.