Replace nested partial_instance function with functools.partial for picklability

[euxhenh]

Description

The current implementation of adapt_class_type uses a locally defined function partial_instance as a wrapper:

https://github.com/omni-us/jsonargparse/blob/dcd0a5ae70d0ef5c7c053436375bcad9aa462c84/jsonargparse/_typehints.py#L1428-L1431

This causes a few issues with pickle serialization, as locally defined functions are not picklable.

Proposed Fix

Replace the custom function with a functools.partial, which is picklable:

return functools.partial(
    instantiator_fn,
    val_class,
    **{**init_args, **dict_kwargs},
)

From what I see this should maintain the same behavior but ensures compatibility with serialization libraries. Can you think of any reason why not to support this?

[mauvilsa]

Sounds good to me. Things being pickleable is better. The only detail is that I don't see *args in the proposed change.

Please go ahead and create a pull request. But also please include a unit test that would fail with the current implementation.

[euxhenh]

@mauvilsa Thanks for the quick reply! functools.partial should handle the *args case similarly - it preserves positional arguments passed later when the partial is called. I.e., the following should work

f = functools.partial(
    instantiator_fn,
    val_class,
    **{**init_args, **dict_kwargs},
)

This can be called as:

f(*args)  # same as instantiator_fn(val_class, *args, **{**init_args, **dict_kwargs})

I'll work on a pull request.

[mauvilsa]

Done in #728