microsoft / microxcaling

PyTorch emulation library for Microscaling (MX)-compatible data formats
MIT License
123 stars 14 forks source link

Consider a PyPI package? #22

Open awf opened 1 month ago

awf commented 1 month ago

Our package https://github.com/graphcore-research/gfloat would like to depend on mx, but currently has to include it in requirements.txt as a direct dependency:

# Requirements for tests
pytest
ml_dtypes
mx @ git+https://github.com/microsoft/microxcaling

Normally, I think this is a good idea - one knows exactly what code one is depending on, etc.

However, pypi says

Can't have direct dependency: mx@ git+https://github.com/microsoft/microxcaling ; 
extra == "dev". See https://packaging.python.org/specifications/core-metadata for 
more information.

And this seems hard to fix (see the "amusing side effect" at https://stackoverflow.com/a/54894359)

If microxscaling could be a PyPI package we would not need the direct dependency.

mgolub2-ms commented 1 month ago

Hi @awf !

Really glad to see you are using our library!

I believe we can make that happen, though I'll need to ping some people to get the right checkboxes checked. Will get back to you ASAP!