argparse.ArgumentParser instances are not pickleable

[mauvilsa]

Bug report

Bug description:

argparse.ArgumentParser instances are not pickleable, and being pickleable is necessary for multiprocessing to work. For example pytorch-lightning sends ArgumentParser objects between processes.

Currently the following fails:

import argparse, pickle
parser = argparse.ArgumentParser()
pickle.loads(pickle.dumps(parser))

giving error:

_pickle.PicklingError: Can't pickle local object <function ArgumentParser.__init__.<locals>.identity

There is the following workaround that is being used in jsonargparse:

import argparse, pickle

def _identity(value):
    return value

class ArgumentParser(argparse.ArgumentParser):
    def __init__(self, *args, **kwargs) -> None:
        super().__init__(*args, **kwargs)
        self.register('type', None, _identity)

parser = ArgumentParser()
pickle.loads(pickle.dumps(parser))

Even though the workaround solves the issue, there is no guarantee that changes in argparse make parsers unpickleable due to other reasons. For instance on Python 3.14.3 we are seeing the error:

_pickle.PicklingError: Can't pickle local object <function HelpFormatter._set_color.<locals>.<lambda>

It would be preferable that ArgumentParser is guaranteed to be pickleable here in cpython.

CPython versions tested on:

3.9, 3.10, 3.11, 3.12, 3.13, 3.14, main

Operating systems tested on:

Linux

Linked PRs

• gh-144782: Make sure that ArgumentParser instances are pickleable #144783

[mpkocher]

I think this might run into challenges depending on how people write the parser instance setup or "glue" around it.

Is this supposed to work?

p = ArgumentParser()
p.add_argument('-n', '--name', type=lambda x: x.upper(), required=True)
pickle.dumps(p)

Perhaps the argparse part might improved, but how people hook in to argparse, or put the glue around It will be fragile.

For example, I suspect patterns or styles like this will never work (and shouldn't be expected to be pickleable).

[mauvilsa]

This issue doesn't suggest that all possible parsers should be pickleable. But it should be possible to build parsers with the intention of being pickleable.

A parser that is not pickleable because of a lambda in an add_argument is not a blocker. The problematic add_argument can be fixed in each project if needed. However, argparse.ArgumentParser itself not being pickleable, e.g. due to the new _set_color.<locals>.<lambda>, is a blocker. Multiprocessing in pytorch-lightning has been working for years without any issue thanks to the self.register('type', None, _identity) workaround. But now with Python 3.14.3 this is broken.