Exception using dataclasses with v1.9

[EthanAtLL]

Running uv run doctest.py on

# /// script
# dependencies = [
#   "numpydoc==1.8.0",
# ]
# ///
from dataclasses import dataclass
from numpydoc.validate import validate, get_doc_object

@dataclass
class MyClass:
    """Docs.
    """
    x: int

if __name__=='__main__':
    doc_obj = get_doc_object(MyClass)
    x = get_doc_object(MyClass.__init__)
    print(validate(x))

works yielding {'type': 'function', 'docstring': '', 'deprecated': False, 'file': '<string>', 'file_line': None, 'errors': [('GL08', 'The object does not have a docstring')], 'examples_errors': ''} but changing to numpydoc==1.90 gives

Traceback (most recent call last):
  File ".../doctest.py", line 18, in <module>
    print(validate(x))
  File ".../uv/environments-v2/doctest-76a1dd0b6c7db47d/lib/python3.9/site-packages/numpydoc/validate.py", line 648, in validate
    cls = Validator._load_obj(f"{doc.code_obj.__module__}.{cls_name}")
  File ".../uv/environments-v2/doctest-76a1dd0b6c7db47d/lib/python3.9/site-packages/numpydoc/validate.py", line 268, in _load_obj
    obj = getattr(obj, part)
AttributeError: module '__main__' has no attribute '__create_fn__'

The error only arises with the @dataclass annotation.

[DerWeh]

I have the same issue, in case further logs help you:
https://app.readthedocs.org/projects/gftools/builds/28675578/
https://github.com/DerWeh/gftools/actions/runs/15946217959/job/44980605108?pr=45

[mattgebert]

@DerWeh it seems your bug is the same as #638, but distinct from this one.

@EthanAtLL I've gone through the code the this issue arises from the use of the code_obj to resolve the parent class (via __qualname__ attribute).

__qualname__ was introduced in Python 3.3. In the dataclasses.py module, the processing path for adding the __qualname__ for __init__ is

dataclasses.py > _process_class > _set_new_attribute(cls, '__init__', init_fn( ... ) )

Here the _set_new_attribute is responsible for binding the __init__ function to the new class.

It looks like in Python 3.9, there is no support in the dataclasses [Py3.9] library for setting __qualname__ for the created instances/functions by _set_new_attribute - the only reference to __qualname__ is in creating a __repr__ function.

In Python 3.10, many __qualname__ modificaitons are introduced to dataclasses [Py3.10], including all uses of the _set_new_attribute method.

@rossbar I'm not familiar enough with this sort of bug to know if we ought to introduce a python version specific fix. Python 3.9 end of life is Oct 2025 as well, if relevant? This issue is very much localized to dataclasses.

[larsoner]

My vote would be to require Python 3.10+ for numpydoc. It's maybe a few months earlier than we would have preferred to make the jump but seems reasonable. People on 3.9 and older can use older versions if needed

[EthanAtLL]

@mattgebert Thank you for looking into this! @larsoner With 3.9 EOLing soon, requiring 3.10+ makes sense.