This project has a very light NumPy and no CuPy dependency, so pinning was briefly useful for the transition when NumPy 2 came out, but unnecessary as soon as there were no obvious Python incompatibilities.
(The dependency is so light, there really should be no issue at all.)

@jakirkham I am a bit confused by the result here. The tests seem to pick up NumPy 2.1 (good), but are also forced to use pytorch 2.0.0. However, pytorch==2.0.0 should not be compatible with NumPy 2?

Considering how simple the dependency on NumPy is, I don't have a concern at all here (that is a torch/NumPy problem), but it does confuse me right now.
However, if pytorch==2.0.0 is required here, it may even make sense to keep the NumPy pin in light of the torch limitation.

EDIT: To clarify this confusion. pytorch seems installed in a pretty aggressive way, bypassing the incompatible versions. But torch is also somewhat forgiving about not requiring numpy for everything, so it seems that things just work out for the light usage here.

Realized that wholegraph is needed for cugraph. Since usage here is increadibly light, I would suggest putting it in soon, to see how the cugraph CI itself does.

Sebastian and I discussed the PyTorch question offline

We observed a few things about PyTorch usage here:

1. PyTorch appears to be a dependency, but is not listed as one
2. Old versions of PyTorch are being installed in CI

These are documented in issue: #210

This led to a separate question. PyTorch only started supporting NumPy 2 in PyTorch 2.3.0 ( pytorch/pytorch#107302 (comment) ). However the versions here are older than that. We observed:

1. Older PyTorch installed with NumPy 2 seems to be working here on CI
2. Testing an older PyTorch locally import torch works with NumPy 2
3. Testing tensor creation in PyTorch locally only issues a warning (no error)

While it doesn't appear necessary to change PyTorch here, we still think it is a good idea to add the PyTorch dependency and update the versions used. However we don't know enough about how PyTorch is used here to develop a solution without more discussion. To facilitate discussion and reaching an acceptable solution, have raised issue: #210

One job had trouble finding the libraft package. While it doesn't explicitly show a network error, it is quit likely this happened. The other jobs in that group found libraft without issues

Restarted that job and it appears to have cleared up

So just noting this for posterity

/merge

Thanks Sebastian! 🙏