Test against PyTorch? CuPy?

[mdhaber]

I thought it would be sufficient to test against strict and numpy, but there may be value in testing against PyTorch and maybe even CuPy. Perhaps the data issue we just fixed is unique, but it's hard to tell without trying. At the very least, it will expose shortcomings in either array_api_compat, array_api_strict, or the Array API Standard documentation). For instance, I immediately ran into data-apis/array-api-compat#226 today, and also found:

# These all work with strict
# import array_api_strict as xp
from array_api_compat import torch as xp
x = xp.asarray(1, dtype=xp.uint32)
-x  # RuntimeError: "neg_cpu" not implemented for 'UInt32'
abs(x)  # RuntimeError: "abs_cpu" not implemented for 'UInt32'
~x  # RuntimeError: "bitwise_not_cpu" not implemented for 'UInt32'

@lucascolley any chance you'd be interested in taking this on? I ask because you've worked on getting SciPy tests running on other arrays.

[lucascolley]

yeah, we have some infrastructure for doing this in https://github.com/data-apis/array-api-extra/blob/main/tests/conftest.py and https://github.com/data-apis/array-api-extra/blob/main/src/array_api_extra/_lib/_backends.py. I agree that it is useful, and have the same planned for quantity-array.

Pretty busy at the moment so not sure when I'll get round to it, but I could probably have a go at tackling it at some point.