Adding cupy support

[JoOkuma]

Hello, first of all, thanks for the amazing package.

Any thoughts on adding cupy support?
I think it should be straightforward since they already support DLpack.

I can implement it, given some pointers.

If it's already supported, I might need some help 😅.
I tried passing a cupy.ndarray and DLpack Capsule using ndarray with nb::device::cuda and it didn't work.

[wjakob]

Does cupy.ndarray have an __dlpack__ member? If so, I'm surprised that it does not work. Otherwise, what's the way of extracting the DLPack capsule?

[JoOkuma]

Yes, they do.

My impression is that nanobind type checks and fails before reaching the __dlpack__ call.

E           TypeError: _cu_func(): incompatible function arguments. The following argument types are supported:
E               1. _cu_func(arg0: ndarray[dtype=uint32, shape=(*, *), order='C', device='cuda'], /) -> numpy.ndarray[dtype=uint64]
E    
E           Invoked with types: cupy.ndarray

[JoOkuma]

Hi @wjakob, any additional thoughts on this?

[wjakob]

I don't use CuPy myself but am open to changes. Would you be able to create a PR that extends the framework detection to handle this case?

[JoOkuma]

Sure, I created #594. Let me know if I missed something, thanks!