where not following Python scalar promotion rules?

[mdhaber]

According to the standard for where,

array_api_strict.where does not seem to implement this rule.

import array_api_strict as xp
x = xp.asarray([1., 2.])
xp.where(x > 1.5, x, 0)
# TypeError: array_api_strict.float64 and array_api_strict.int64 cannot be type promoted together

[ev-br]

So the relevant wording from the spec is:

Using Python scalars (i.e., instances of bool, int, float, complex) together with arrays must be supported for:

array <op> scalar


... and scalar has a type and value compatible with the array data type:


a Python int or float for real-valued floating-point array data types.

so indeed where is too strict here. PR welcome!

[ev-br]

What I do wonder though is that array <op> scalar is more permissive than array <op> array:

In [9]: x = xp.asarray(1.)

In [10]: x + 0
Out[10]: Array(1., dtype=array_api_strict.float64)

In [11]: x + xp.asarray(0)
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
...
TypeError: array_api_strict.float64 and array_api_strict.int64 cannot be type promoted together

This was definitely discussed and rejected (at a guess, it's granularity of floating point numbers gets larger than 1 at the end for values >> 1). Anyhow, it's covered by "A conforming implementation of the array API standard may support additional type promotion rules beyond those described in this specification."

[ev-br]

PR welcome!

Well it's easiest to just fix it, now that I very recently worked on a similar support for other binary functions, and still remember the details.

gh-132 seems to do the trick.

[mdhaber]

I also ran into:

import array_api_strict as xp
xp.where(xp.asarray([True, False]), 1., xp.asarray([2, 2], dtype=xp.float32)
Array([1., 2.], dtype=array_api_strict.float64)  # should be float32

Will gh-132 also fix this?
Is there reason to believe the problem is isolated to where? I tested clip (very lightly) and it seems OK, and I notice that where was a different PR from the rest.

[ev-br]

Will #132 also fix this?

Yes. It basically makes where reuse scalar handling from binary functions.

In [1]: import array_api_strict as xp

In [2]: xp.where(xp.asarray([True, False]), 1., xp.asarray([2, 2], dtype=xp.float32))
Out[2]: Array([1., 2.], dtype=array_api_strict.float32)

In [3]: xp.__version__
Out[3]: '2.4.dev3+g33b4bf6.d20250302'

[mdhaber]

Thanks. I think I am able to remove all the test skips related to this issue now.