Lazy imports break dataclass InitVars in extremely rare scenarios #160665

Bug report

3 files are required to get this bug to happen:

testing.py

from testing_2 import A
from testing_3 import B

a = A()

b = B(a)

testing_2.py

from dataclasses import dataclass

from testing_3 import B


@dataclass
class A:
    pass

testing_3.py

from dataclasses import InitVar, dataclass

lazy from testing_2 import A


@dataclass
class B:
    a: InitVar[A]

    def __post_init__(self, a: A) -> None:
        print(a)

If you run testing.py with python 3.15 you will get TypeError: B.__post_init__() missing 1 required positional argument: 'a'
The key thing that causes the error seems to be the presence of the import cycle combined with the lazy import. It's worth noting that if you use a TYPE_CHECKING guard instead as described below, it does also work on 3.15.

I found this issue while migrating a bunch of my TYPE_CHECKING guards to lazy imports - infact as PEP 810 says "Lazy imports eliminate the common need for TYPE_CHECKING guards". The version of python that breaks is cpython-3.15.0a7-linux-x86_64-gnu. The code instead works on cpython-3.14.3-linux-x86_64-gnu, if of course you replace the lazy import in testing_3.py with a if TYPE_CHECKING: block

CPython versions tested on:

3.15

Operating systems tested on:

Linux

Linked PRs

• gh-160665: Recognize dataclass markers deferred by lazy imports #160668