Charles Frye

In the conversion from 4 to 5, Jupyter added a "feature" that halts data transmission from the kenrel to the client if it exceeds a certain threshold. This was intended...

as on e.g. 67b848c192d8ae232f7633fbefc3922b0e077b47

[This paper](https://transformer-circuits.pub/2021/framework/index.html) provides a mathematical framework for thinking about attention-only transformers. If we drop the softmax, this becomes a pretty solid demo for Transformers in `panndas` -- borrowing the details...

we should pin model revisions on Hugging Face

1

models can still be yanked, but this should reduce the variability Here's an example, which uses tgi: ```py def download_model(): subprocess.run( [ "text-generation-server", "download-weights", MODEL_ID, "--revision", REVISION, ], env={ **os.environ,...