Parsing `TYPE_CHECKING` guarded import doesn't work in combination with `Type`

[EthanMarx]

🐛 Bug report

When using the Type annotation in conjunction with a TYPE_CHECKING guarded import, e.g.

from typing import Type, TYPE_CHECKING
from jsonargparse import ArgumentParser

if TYPE_CHECKING:
    from pytorch.lightning.cli import LightningCLI

def dummy(cli_cls: Type["LightningCLI"]):
    print(cli_cls)


def main():
    parser = ArgumentParser()
    parser.add_function_arguments(dummy)
    args = parser.parse_args()

I get the following error

ValueError: With fail_untyped=True, all mandatory parameters must have a supported type. Parameter 'cli_cls' from '__main__.test' does not specify a type.

Although, this parser set up works when removing the TYPE_CHECKING guarded import, and importing directly

Expected behavior

Environment

• jsonargparse version: 4.3.1
• Python version: 3.9
• How jsonargparse was installed: poetry
• OS: linux

[mauvilsa]

Thank you for reporting! Would you be interested in contributing the fix?

[EthanMarx]

Sure - would you be able to point me in the right direction?

[mauvilsa]

Certainly.

1. Setup a local environment as explained in CONTRIBUTING.rst#development-environment.
2. The tests related to TYPE_CHECKING are in jsonargparse_tests/test_postponed_annotations.py, so make sure those tests run, like

   pytest -v jsonargparse_tests/test_postponed_annotations.py
   

3. Have a look at the existing tests: test_postponed_annotations.py#L192-L294
4. Maybe, debug through one of the tests to get a bit of understanding of how things work.
5. Before trying to fix, extend test_postponed_annotations.py to include a test for this case which is failing. That is, add an entry in the if TYPE_CHECKING: block and implement a test_* function that fails as you have observed with the current version of the code.

The unit test would already be a valuable contribution. The following step would be to understand what is currently happening and figure out where to fix/improve the code. This feature makes use of AST parsing, so it might not be trivial to understand and fix. But you can give it a shot. If it seems to complex just ask again, or I could continue from where you left off.

[EthanMarx]

Thank you! I actually think I found a solution. I have a PR with unit tests ready to go, but can't get past the pre-push testing hooks, I think due to #574

[mauvilsa]

You can do git push --no-verify. Or push from an env that doesn't have pytorch installed.