Allow for str versions of dicts based on typing

[muellerzr]

What does this PR do?

This PR adds support for passing in a string dictionary to arguments that allow a dict in the argparser. For example:

python test.py --output_dir testdir --accelerator_config='{"dispatch_batches":"False"}'

(test.py is just a small script reading the args from the parsers):

from transformers import HfArgumentParser, TrainingArguments

parser = HfArgumentParser((TrainingArguments))

training_args = parser.parse_args_into_dataclasses()

print(training_args)

This also fixes issues with typing of not being able to state that deepspeed and fsdp_config can't have type dict. Turns out it's a (very) fun setting with the argparser wrapper we have for these. Types should be declared as (for this):

Optional[Union[dict,str,...]]

This is very important to get working properly!

Fixes #30204

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@amyeroberts @pacman100

Technically everything will go brr if passes, since any call to the CLI/parser happens usually with examples etc. So new tests added, but did manually verify input/output from a basic script locally

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[muellerzr]

@amyeroberts figured out where to add a test, and verified that all 3 conditions raise their respective errors :)

[amyeroberts]

Thanks for enabling this!

[amyeroberts]

I like just checking for a few possible args!

[amyeroberts]

I know on offline discussion I said we probably don't need a test to check the parsing. Seeing the implementation, i.e. TrainingArguments can be created with a field as a string representation of a dict, and not having to include the CLI, I think we can add a simple for at least one of the fields in VALID_DICT_FIELDS

e.g. something along the lines of:

def test_valid_dict_input_parsing(self):
    args = TrainingArguments(
        field_name='{"key": value}'
    )
    # Or however it's assigned in the args
    self.assertEqual(args.field_name, {key: value})

[muellerzr]

Done. This also made me notice that we cast int and bools as str still, so added a helper for this (and does so to avoid literal_eval, which can be exploited)

[pacman100]

Thank you @muellerzr for adding this super useful feature to be able to pass string representation of dict as cmd arguments! 🚀

This should be very useful for the CLI support that @younesbelkada worked wrt TRL.

[amyeroberts]

Great! Thanks for enabling this and adding these tests ❤️

[amyeroberts]

Nice :)

[amyeroberts]

Nice!