Fix PreTrainedConfig as Pydantic field type after dataclass conversion

[joaquinhuigomez]

Root cause

The v5.4.0 release converted PreTrainedConfig from a regular class to a @dataclass. This changes how Pydantic handles it when used as a field type in a BaseModel: instead of treating it as an opaque arbitrary type, Pydantic now introspects the dataclass fields and attempts to resolve all type annotations.

The dtype field is annotated as Union[str, "torch.dtype"] | None, where "torch.dtype" is a forward reference that depends on torch being in the namespace. Since torch is only imported under TYPE_CHECKING, Pydantic raises PydanticUndefinedAnnotation: name 'torch' is not defined.

Fix

Add __get_pydantic_core_schema__ to PreTrainedConfig that returns a core_schema.is_instance_schema(cls). This tells Pydantic to validate values by type-checking rather than introspecting the dataclass fields, restoring the pre-v5.4.0 behavior.

The method is inherited by all subclasses (e.g. BertConfig), so they also work as Pydantic field types.

Reproduction

from pydantic import BaseModel, ConfigDict, Field
from transformers import PretrainedConfig

class MyModelConfig(BaseModel):
    model_config = ConfigDict(arbitrary_types_allowed=True)
    sub_config: PretrainedConfig = Field(description="Configuration for the sub_config")

MyModelConfig.model_rebuild(force=True)  # raises PydanticSchemaGenerationError on v5.4.0

After this fix, the above passes.

Fixes #45070

[github-actions]

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=45080&sha=7aa9f5

[joaquinhuigomez]

Going to close this — needs more investigation than I initially thought.