scipy-stubs: `scipy.optimize.least_squares` function inferred as module

[jorenham]

As reported in scipy/scipy-stubs#939, mypy seems to infer scipy.optimize.least_squares as a ModuleType. Other type-checkers (pyright, basedpyright, pyrefly) correctly infer it as a function.

from typing import reveal_type
from scipy.optimize import least_squares

reveal_type(least_squares)

• mypy: Revealed type is "types.ModuleType"
• pyrefly: revealed type: (fun: (🍝) -> OptimizeResult
• (based)pyright: Type of "least_squares" is "(fun: (🍝) -> OptimizeResult"

(🍝 = the huge function signature)

This seems to occur on (at least) scipy-stubs 1.16.2.0 and 1.16.2.3, using mypy 1.18.2.

To repro, install scipy-stubs[scipy]==1.16.2.3 (currently the latest) from pypi or scipy-typed from conda.

Here's are the relevant bits of code:

1. optimize/__init__.pyi re-exports from ._lsq import least_squares through __all__
2. optimize/_lsq/__init__.pyi re-exports from .least_squares import least_squares through __all__
3. optimize/_lsq/least_squares.pyi contains the definition for the least_squares function

It looks like mypy ignores step optimize/_lsq/__init__.pyi from 2., causing it to resolve the optimize._lsq.least_squares module, instead of (explicitly re-exported) optimize._lsq.least_squares.least_squares function.

I expect to be able to work around this by changing from ._lsq import least_squares in optimize/__init__.pyi to from ._lsq.least_squares import least_squares.

[jorenham]

Oh btw, in case it wasn't clear, mypy's behavior here contradicts the runtime semantics:

>>> from scipy.optimize import least_squares
>>> type(least_squares)
<class 'function'>
>>> from scipy.optimize._lsq import least_squares
>>> type(least_squares)  # a module according to mypy
<class 'function'>
>>> from scipy.optimize._lsq.least_squares import least_squares
>>> type(least_squares)
<class 'function'>

[ilevkivskyi]

Hm, does it help if you write

from .least_squares import least_squares as least_squares

?

[jorenham]

Hm, does it help if you write

from .least_squares import least_squares as least_squares
?

Hmm no, I see no difference when I change scipy-stubs/optimize/_lsq/__init__.pyi into this:

from .least_squares import least_squares as least_squares
from .lsq_linear import lsq_linear as lsq_linear

__all__ = ["least_squares", "lsq_linear"]

removing the cache didn't help either.

And it's the same with or without the __all__

[randolf-scholz]

Probably because least_squares is both a submodule _lsq/least_squares.py and a member of _lsq/__init__.py; so from scipy.optimize._lsq import least_squares is ambiguous: are we importing the submodule or the identically names attribute of _lsq/__init__.py ? Seems that runtime prefers the member of __init__.py in that case, so mypy needs to do the same.