ENH: allow tuple pad_width in pad

[ev-br]

Observed in scipy/scipy#22226 : allow pad_width to be a tuple.

[ev-br]

Testing scipy/scipy#22122 and scipy/scipy#22226 shows the need to bite the bullet and add pad_width being an int or a 2-tuple or a list of 2-tuples.
When round-tripping through numpy, it was all hidden out of sight; otherwise, need to implement this manually, as this PR does now.

[lucascolley]

The remaining mypy failures can be ignored with inline comments, I think.

[ev-br]

I really don't understand what on earth it wants and how to silence it. Now taht you want all this on, I'd appreciate if you could silence it with whatever means you prefer as appropriate.
(in my projects, I'd just shut all this off and be done with it; you clearly prefer some other way)

[lucascolley]

will do!

[lucascolley]

@DetachHead Basedpyright seems to be missing this assignment on line 599. With or without the explicit cast I've added here, we see:

/home/runner/work/array-api-extra/array-api-extra/src/array_api_extra/_funcs.py:609:32 - error: Argument of type "int | tuple[int, int] | list[tuple[int, int]]" cannot be assigned to parameter "iterable" of type "Iterable[_T@enumerate]" in function "__new__"
  Type "int | tuple[int, int] | list[tuple[int, int]]" is not assignable to type "Iterable[_T@enumerate]"
    "int" is incompatible with protocol "Iterable[_T@enumerate]"
      "__iter__" is not present (reportArgumentType)

https://github.com/data-apis/array-api-extra/actions/runs/12651016113/job/35250714798?pr=82

By line 609 pad_width must be a list. Is this a known issue?

[lucascolley]

adding the three casts in

diff --git a/src/array_api_extra/_funcs.py b/src/array_api_extra/_funcs.py
index d5ee7fe..8fa397b 100644
--- a/src/array_api_extra/_funcs.py
+++ b/src/array_api_extra/_funcs.py
@@ -596,7 +596,7 @@ def pad(
 
     # make pad_width a list of length-2 tuples of ints
     if isinstance(pad_width, int):
-        pad_width = cast(list[tuple[int, int]], [(pad_width, pad_width)] * x.ndim)
+        pad_width = [(pad_width, pad_width)] * x.ndim
 
     if isinstance(pad_width, tuple):
         pad_width = [pad_width] * x.ndim
@@ -606,6 +606,7 @@ def pad(
 
     slices = []
     newshape = []
+    pad_width = cast(list[tuple[int, int]], pad_width)
     for ax, w_tpl in enumerate(pad_width):
         if len(w_tpl) != 2:
             msg = f"expect a 2-tuple (before, after), got {w_tpl}."
@@ -624,6 +625,8 @@ def pad(
         newshape.append(sh)
         slices.append(sl)
 
+    newshape = cast(list[int], newshape)
+    slices = cast(list[slice], slices)
     padded = xp.full(
         tuple(newshape),
         fill_value=value,
diff --git a/tests/test_funcs.py b/tests/test_funcs.py
index 4ae57ea..2fc9041 100644
--- a/tests/test_funcs.py
+++ b/tests/test_funcs.py
@@ -426,7 +426,7 @@ class TestPad:
         assert padded.shape == (6, 7)
 
         with pytest.raises(ValueError, match="expect a 2-tuple"):
-            pad(a, [(1, 2, 3)])  # type: ignore[list-item]
+            pad(a, [(1, 2, 3)])  # type: ignore[list-item]  # pyright: ignore[reportArgumentType]
 
     def test_list_of_tuples_width(self):
         a = xp.reshape(xp.arange(12), (3, 4))

fixes the basedpyright errors, but now basedmypy complains about a redundant cast:

src/array_api_extra/_funcs.py:609:17: error: Redundant cast to "list[(int, int)]"  [redundant-cast]
        pad_width = cast(list[tuple[int, int]], pad_width)

[lucascolley]

@KotlinIsland looks like there is something basedmypy is catching which basedpyright isn't? Any idea what?

[DetachHead]

it looks like pyright isn't narrowing it because x is an alias to Any. here's a minified repro

from typing import Any


def _(value: tuple[int, ...] | list[int] | int, x: Any):
    if isinstance(value, int):
        value = [value]

    reveal_type(value) # list[int] | tuple[int, ...]
    
    if isinstance(value, tuple):
        value = [1] * x
    
    reveal_type(value) # tuple[int, ...] | list[int] | int

i think this behavior is more correct than what mypy is doing, because the Any could have an __rmul__ that changes the type to something completely different. in fact i think this should probably widen value to Any instead because the type could be changed to anything.

[lucascolley]

Good point, thanks for the fast response! I'll add a cast and ignore the mypy error then 👍

[lucascolley]

the error is reproduced in CI: https://github.com/data-apis/array-api-extra/actions/runs/12654902636/job/35263897195?pr=82

[lucascolley]

this playground has the entire function almost unmodified and it still doesn't reproduce the failure: https://mypy-play.net/?mypy=basedmypy-latest&python=3.13&gist=11d14b9d1c9bd562ef9b6554268e97f5

[KotlinIsland]

the error is that slice takes three generic parameters, i think the playground has an outdated version from prior to this.
ymyzk/mypy-playground#821

i'm not sure why it's reporting explicit any, i would expect it to report "Missing type parameters for generic type"

[lucascolley]

thanks, that makes sense why I can't reproduce on the playground. With the following diff:

diff --git a/src/array_api_extra/_funcs.py b/src/array_api_extra/_funcs.py
index 2f4bc0f..0c9f3e3 100644
--- a/src/array_api_extra/_funcs.py
+++ b/src/array_api_extra/_funcs.py
@@ -605,7 +605,7 @@ def pad(
         xp = array_namespace(x)
 
     # https://github.com/data-apis/array-api-extra/pull/82#discussion_r1905688819
-    slices: list[slice] = []  # type: ignore[no-any-explicit]
+    slices: list[slice[int | None, int | None, None]] = []
     newshape: list[int] = []
     for ax, w_tpl in enumerate(pad_width):
         if len(w_tpl) != 2:
@@ -613,6 +613,7 @@ def pad(
             raise ValueError(msg)
 
         sh = x.shape[ax]
+        sl: slice[int | None, int | None, None]
         if w_tpl[0] == 0 and w_tpl[1] == 0:
             sl = slice(None, None, None)
         else:

I now see:

src/array_api_extra/_funcs.py:618:18: error: Incompatible types in assignment (expression has type "slice[int | Any, int | Any, int | Any]", variable has type "slice[int | None, int | None, None]") 
[assignment]
                sl = slice(None, None, None)
                     ^~~~~~~~~~~~~~~~~~~~~~~
src/array_api_extra/_funcs.py:623:18: error: Incompatible types in assignment (expression has type "slice[int | Any, int | Any, int | Any]", variable has type "slice[int | None, int | None, None]") 
[assignment]
                sl = slice(start, stop, None)
                     ^~~~~~~~~~~~~~~~~~~~~~~~

can you see where the Any is coming from? Perhaps something isn't working with slice being generic over None?

[KotlinIsland]

def __new__(
        cls, start: int | None, stop: int | None, step: int | None = None, /
    ) -> slice[int | MaybeNone, int | MaybeNone, int | MaybeNone]: ...

where MaybeNone = Any

and the reason a: slice contains explicit Any is that:

_StartT = TypeVar("_StartT", covariant=True, default=Any)
_StopT = TypeVar("_StopT", covariant=True, default=Any)
_StepT = TypeVar("_StepT", covariant=True, default=Any)

class slice(Generic[_StartT, _StopT, _StepT]):

therefore:

a: slice
# is the same as
a: slice[Any, Any, Any]

so this is an issue with typeshed: python/typeshed#13376

[lucascolley]

thanks @ev-br, merging!

do let me know if you have a clue what is going on with the failure I can't reproduce in the playground @KotlinIsland ! EDIT: just saw your reply