Hotfix: Add ParallelismConfig fallback for transformers with old accelerate.

Fix TrainingArguments.parallelism_config NameError with accelerate<1.10.1:

• Tests with minimum versions: https://github.com/huggingface/trl/actions/runs/17635808933/job/50111711153

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

E           NameError: name 'ParallelismConfig' is not defined

Problem

When accelerate<1.10.1, the module accelerate.parallelism_config does not exist.
Transformers’ TrainingArguments currently annotates:

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

This creates a ForwardRef. On Python ≥3.12, when TRL constructs its CLI parsers, transformers.HfArgumentParser calls typing.get_type_hints(TrainingArguments). Since ParallelismConfig is undefined in transformers.training_args, a NameError is raised:

NameError: name 'ParallelismConfig' is not defined

Fix

This PR add a compat helper that pre-binds transformers.training_args.ParallelismConfig = Any if the name is missing. This ensures get_type_hints can always resolve the reference. The helper is called early before any HfArgumentParser is created.

Upstream

The proper fix is proposed in:

• Fix TrainingArguments.parallelism_config NameError with accelerate<1.10.1 transformers#40818