Fix TrainingArguments.parallelism_config NameError with accelerate<1.10.1

[albertvillanova]

Fix TrainingArguments.parallelism_config NameError with accelerate<1.10.1:

NameError: name 'ParallelismConfig' is not defined

RuntimeError: Type resolution failed for <class 'transformers.training_args.TrainingArguments'>. Try declaring the class in global scope or removing line of `from __future__ import annotations` which opts in Postponed Evaluation of Annotations (PEP 563)

This PR fixes a runtime compatibility issue in TrainingArguments with accelerate<1.10.1.

Problem

With accelerate<1.10.1, the module accelerate.parallelism_config does not exist.
TrainingArguments currently annotates the field as:

parallelism_config: Optional["ParallelismConfig"] = field(default=None)

When HfArgumentParser calls typing.get_type_hints on this dataclass, Python 3.12 attempts to resolve the ForwardRef("ParallelismConfig") and fails.

See failing CI in downstream TRL: https://github.com/huggingface/trl/actions/runs/17635808933/job/50111711153

Minimal Reproducible Example

pip install transformers accelerate==1.4.0

from transformers import HfArgumentParser, TrainingArguments

# This will trigger get_type_hints on TrainingArguments
parser = HfArgumentParser(TrainingArguments)

Traceback (most recent call last):
  File "huggingface/transformers/src/transformers/hf_argparser.py", line 258, in _add_dataclass_arguments
    type_hints: dict[str, type] = get_type_hints(dtype)
                                  ^^^^^^^^^^^^^^^^^^^^^
  File ".pyenv/versions/3.12.9/lib/python3.12/typing.py", line 2277, in get_type_hints
    value = _eval_type(value, base_globals, base_locals, base.__type_params__)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".pyenv/versions/3.12.9/lib/python3.12/typing.py", line 430, in _eval_type
    ev_args = tuple(
              ^^^^^^
  File ".pyenv/versions/3.12.9/lib/python3.12/typing.py", line 431, in <genexpr>
    _eval_type(
  File ".pyenv/versions/3.12.9/lib/python3.12/typing.py", line 415, in _eval_type
    return t._evaluate(globalns, localns, type_params, recursive_guard=recursive_guard)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".pyenv/versions/3.12.9/lib/python3.12/typing.py", line 947, in _evaluate
    eval(self.__forward_code__, globalns, localns),
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "<string>", line 1, in <module>
NameError: name 'ParallelismConfig' is not defined

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "huggingface/transformers/src/transformers/hf_argparser.py", line 143, in __init__
    self._add_dataclass_arguments(dtype)
  File "huggingface/transformers/src/transformers/hf_argparser.py", line 260, in _add_dataclass_arguments
    raise RuntimeError(
RuntimeError: Type resolution failed for <class 'transformers.training_args.TrainingArguments'>. Try declaring the class in global scope or removing line of `from __future__ import annotations` which opts in Postponed Evaluation of Annotations (PEP 563)

Solution

• Always define the symbol ParallelismConfig at module scope: it is only used as type annotation
  • Use the real class if accelerate>=1.10.1, otherwise fall back to Any.
• Replace the string annotation with Optional[ParallelismConfig] to avoid creating a ForwardRef

This ensures get_type_hints(TrainingArguments) works across Python versions and with older accelerate installs.

[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.

[qgallouedec]

LGTM, waiting for @SunMarc for final review

[SunMarc]

Indeed, thanks for this fix ! cc @S1ro1 for visibility, this created some typing issue for ppl relying on older version of accelerate

[S1ro1]

Thanks for the fix, didn't know this would break type-checkers, my bad.