Allow Python scalars in array_namespace

[asmeurer]

They are just ignored. This makes array_namespace easier to use for functions that accept either arrays or scalars.

I'm not sure if I should have this behavior by default, or if it should be enabled by a flag.

[asmeurer]

We can also automatically ignore None inputs. The scipy helper does this so it would clearly be useful.

[lucascolley]

I would say this should be opt-in, given the standard really focusses on just arrays. In practice, it will probably be used more often than not (at least for now), but given that (for now) most projects will wrap array_namespace, it isn't a problem to have the extra argument.

In SciPy, it doesn't really matter either way as we coerce scalars to NumPy arrays before passing them to array_namespace.

[asmeurer]

There are some active discussions on allowing scalars in more places in the array API. See data-apis/array-api#805 and data-apis/array-api#807.

but given that (for now) most projects will wrap array_namespace, it isn't a problem to have the extra argument.

I'm trying to make that not the case. Some projects like SciPy will have specific reasons they'll need to wrap things, but ideally whatever most projects need should be implemented already in the wrappers here.

[lucascolley]

I think another form of wrapping that will be very common is using a default backend when all arguments of array_namespace are Python scalars.

[asmeurer]

That would be easy to add. I would caution against it though unless you are doing it for backwards compatibility. If you default to NumPy, for instance, you could end up in a situation where a user gets a NumPy array from one function because they passed in scalars, and another array from another function because they passed in arrays.

[lucascolley]

Yeah, I'm thinking that the reason would be to maintain BC

[ev-br]

From experience adding array api support to scipy.ndimage and scipy.signal:

• ignoring Nones is a big +1. As mentioned, scipy has a helper anyway, and it does not seem risky
• ignoring python scalars would be nice actually. scipy has one-off helpers, which basically replace a scalar with an int a None, cf https://github.com/scipy/scipy/blob/main/scipy/ndimage/_delegators.py#L201 ; This is not a deal-breaker though, because scipy needs other helpers, too (_skip_if_dtype etc).
• adding logic of the sort "assume numpy if all scalars" does not look desirable. "In the face of ambiguity" and all that.

In general, I'd say the logic for unknown arguments (None, possibly scalars) should be "ignore an unknown object and deduce the namespace from all other arguments". Scipy does it in a couple of places, e.g. there's a scipy PR which makes xp_assert_close(array, list) convert the list to the namespace of array.

In that view, ignoring scalars is fine as long as array_namespace() raises on no inputs:

In [5]: xp.array_namespace()
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Input In [5], in <cell line: 1>()
----> 1 xp.array_namespace()

File ~/repos/array-api-compat/array_api_compat/common/_helpers.py:563, in array_namespace(api_version, use_compat, *xs)
    560         raise TypeError(f"{type(x).__name__} is not a supported array type")
    562 if not namespaces:
--> 563     raise TypeError("Unrecognized array input")
    565 if len(namespaces) != 1:
    566     raise TypeError(f"Multiple namespaces for array inputs: {namespaces}")

TypeError: Unrecognized array input

[asmeurer]

That's very useful feedback. It sounds like the behavior in this PR matches what you are suggesting (it does indeed raise if inputs are only scalars or None, as you can see from the tests).

Copilot reviewed 2 out of 2 changed files in this pull request and generated no suggestions.

[ev-br]

Tried it locally and it does work as expected. Would be great to be able to use this downstream. So, for what it's worth, LGTM.

[asmeurer]

Looks like this broke some of the tests. We should have merged the latest main before merging here since it was an older PR.

[asmeurer]

This breaks NumPy scalar support because of this annoying fact:

>>> isinstance(np.float64(0), float)
True