Mix & match behavior of array objects from different libraries

[leofang]

This question was raised by @shwina during an internal discussion: If I create two array objects, one from cupy.array_api and another from numpy.array_api, if I do a+b or b+a what would happen? I think we should obviously return TypeError following Python's convention, so the question is which page should we document this expectation?

[rgommers]

In https://data-apis.org/array-api/latest/purpose_and_scope.html#out-of-scope it says:

Non-goals for the API standard include: ... Making it possible to mix multiple array libraries in function calls. Most array libraries do not know about other libraries, and the functions they implement may try to convert “foreign” input, or raise an exception. This behaviour is hard to specify; ensuring only a single array type is used is best left to the end user.

We have tended to not specify exception types and other such things that are out of scope, undefined behavior, etc. TypeError seems clearly the correct exception type, but I'm not sure it's a good idea to document this under the Python operators section or elsewhere. Unless we can think of a good way to have a self-contained section where we document the most common exceptions and when to use them somehow.

[asmeurer]

We don't really require type checks of any kind anywhere. We have type hints, but no requirement that those be translated to checks at runtime. My reading of the spec is that such a thing is garbage in, garbage out. It might produce an error or it might work, and if it does work, it might be on accident. Requiring a + b to raise a TypeError when b is not an array object of the same library as a (or a Python scalar) means a.__add__ has to have a if not isinstance(other, array): raise TypeError check at the top. And the same for every other method and function. Even the strict NumPy implementation does not do this (in fact, numpy.array_api.Array operator methods return NotImplemented, meaning another library could interface with them if it wanted to).

[rgommers]

I believe the question is answered and we already have text that addresses the point, quoted in italics above. So I'll close this issue.