Type of variable different when instantiated in return

[haarisr]

import numpy as np
import numpy.typing as npt


def foo() -> npt.NDArray[np.float64]:
    return np.array([1.0, 2.0, 3.0], dtype=np.float32)

def bar() -> npt.NDArray[np.float64]:
    temp = np.array([1.0, 2.0, 3.0], dtype=np.float32)
    return temp

try.py:10:12 - error: Expression of type "NDArray[floating[_32Bit]]" is incompatible with return type "NDArray[float64]"
    "ndarray[Any, dtype[floating[_32Bit]]]" is incompatible with "ndarray[Any, dtype[float64]]"
      Type parameter "_DType_co@ndarray" is covariant, but "dtype[floating[_32Bit]]" is not a subtype of "dtype[float64]"
        "dtype[floating[_32Bit]]" is incompatible with "dtype[float64]"
          Type parameter "_DTypeScalar_co@dtype" is covariant, but "floating[_32Bit]" is not a subtype of "float64"
            "floating[_32Bit]" is incompatible with "floating[_64Bit]" (reportReturnType)
1 error, 0 warnings, 0 informations

I am not sure why foo does not return an error but bar does. I believe foo should also throw an error? I could not create an example with standard types, not sure if it is a numpy issue

[erictraut]

I'm not able to repro the behavior you're seeing. I'm guessing that you're using an older version of numpy. Please try upgrading to the latest version (pip install numpy --upgrade). With the latest version, I'm seeing the return statements in both foo and bar result in type violation errors.

The big difference between foo and bar is that the call to np.array is evaluated with normal inference rules in bar but with bidirectional type inference rules in foo. With bidirectional type inference, pyright takes into account the "expected type", which in this case is provided by the return type annotation.

[erictraut]

Perhaps you have an older type stub library for numpy installed in your python environment? You could confirm by creating a fresh venv and installing numpy.

[haarisr]

uv venv venv -p 3.12 && \
source venv/bin/activate && \
uv pip install pyright numpy --no-cache && \
uv pip show pyright && \
uv pip show numpy && \
pyright try.py && \
cat try.py

Using Python 3.12.4 interpreter at: /usr/bin/python3.12
Creating virtualenv at: venv
Activate with: source venv/bin/activate
Resolved 3 packages in 312ms
Prepared 3 packages in 1.80s
Installed 3 packages in 4ms
 + nodeenv==1.9.1
 + numpy==2.0.1
 + pyright==1.1.373
Name: pyright
Version: 1.1.373
Location: foo/venv/lib/python3.12/site-packages
Requires: nodeenv
Required-by:
Name: numpy
Version: 2.0.1
Location: foo/venv/lib/python3.12/site-packages
Requires:
Required-by:
0 errors, 0 warnings, 0 informations
import numpy as np
import numpy.typing as npt


def foo() -> npt.NDArray[np.float64]:
    return np.array([1.0, 2.0, 3.0], dtype=np.float32)

Still the same

[haarisr]

import numpy as np
import numpy.typing as npt


def foo() -> npt.NDArray[np.float64]:
    return (arr:=np.array([1.0, 2.0, 3.0], dtype=np.float32), reveal_type(arr))

I get the error with this but this might be similar to the bar example again. Without reveal_type there is no error

[erictraut]

Ah, my bad. I'm running the latest (prerelease) version of pyright. This turns out to be a bug that I fixed a couple of days ago (#8519). It will be addressed in this week's release.

[haarisr]

Thanks for the help and the work on pyright!