The pyinstaller error looks unrelated. I wonder if the new version 5.7.0 released Dec 4 is the cause.

EDIT: Some of the passing PRs also use the 5.7.0 version.

The error is

Traceback (most recent call last):
  File "numpy/_pyinstaller/pyinstaller-smoke.py", line 14, in <module>
  File "PyInstaller/loader/pyimod02_importers.py", line 499, in exec_module
  File "numpy/__init__.py", line 143, in <module>
  File "PyInstaller/loader/pyimod02_importers.py", line 499, in exec_module
  File "numpy/core/__init__.py", line 100, in <module>
  File "PyInstaller/loader/pyimod02_importers.py", line 499, in exec_module
  File "numpy/core/_add_newdocs.py", line 4970, in <module>
  File "numpy/core/function_base.py", line 529, in add_newdoc
ModuleNotFoundError: No module named 'numpy.core._multiarray_tests'

and seems to be related to this line

It seems raises and run_module_suite are main names available through np.testing. that are removed. We could add a brief:

def __getattr__(self, name):
    # These were removed in 1.25
    if name == "raises" or name == "run_module_suite":
        raise AttributeError(
            "numpy.testing.raises and numpy.random.run_module_suite wrapped "
            "`nose` and has been removed. We recommend switching to `pytest`.")

But, I am happy either way. The typing test looks like the test for the decorator typing itself.

Maybe the error is due to a circular import somehow when using the pyinstaller hook?

There is a nose "excludedimports" in pyinstaller we can delete, but that seems like a tiny cleanup that shouldn't cause a failure. I restarted the test out of curiosity, but no good idea, the hook seems to collect all dynamic libs in numpy, so missing the module seems strange.

Hmm... the noseclasses file imported from _multiarray_tests, maybe we need another import somewhere.

Importing _multiarray_tests in _add_newdocs.py fixes the missing module problem. The question is if that is where we want to import it.

There are a number of profiling errors of the form:

2023-01-19T21:08:46.8102125Z libgcov profiling error:/home/runner/work/numpy/numpy/build/temp.linux-x86_64-3.9/numpy/core/src/common/npy_hashtable.gcda:overwriting an existing profile data with a different timestamp

Which I don't think should be related unless removing nose has exposed a potential profiling problem.

Ohhh, the tests are the issue... I suspect you can also add it to:

hiddenimports = ['numpy.core._dtype_ctypes']

in the pyinstaller hook file. Not sure about the other thing, but I doubt it is related.

Ohhh, the tests are the issue... I suspect you can also add it to:

I considered that, made me wonder if pyinstaller was working correctly before. In the absence of complaints I decided to leave it as is, especially as I don't really understand it :)

EDIT1: but maybe "hidden" refers to the underscore prefix? But then I would expect problems with the other underscored modules, but they are all imported up top.

EDIT2: I note that numpy.core._dtype_ctypes is an ordinary Python module, but it imports both ctypes and _ctypes which seem to be standard Python libraries.

We should figure out how to deprecate ModuleDeprecationWarning. I wonder if deprecation warnings work properly in exceptions.

Squashed and release note added.

My guess is that things would work fine if the _pyinstaller directory were moved down below conftest.py, or even if the import were added to _pytesttester.py.

Thanks @charris