TRF055

config on a PreTrainedModel subclass must be an annotation (config: SomeConfig), not an assignment (config = SomeConfig).

Default	Enabled
Scope	All models
Source	mlinter/trf055.py
Show in terminal	mlinter --rule TRF055

What it does

Checks that PreTrainedModel subclasses in modeling_*.py and modular_*.py do not assign config = SomeConfig as a class attribute.

Why is this bad?

PreTrainedModel.__init_subclass__ derives config_class by looking for a config **annotation** via inspect.get_annotations(cls). An assignment (config = SomeConfig) is invisible to this mechanism: it creates a stray class attribute but inspect.get_annotations returns None for it, so the subclass falls back to inheriting the parent’s config_class instead of picking up the intended config. A pure annotation (config: SomeConfig) has no runtime value and does not create an attribute, so it is correctly detected by inspect.get_annotations and sets config_class to the right class.

Example

 class Gemma4VisionModel(Gemma4PreTrainedModel):
     """The Gemma 4 Vision Encoder."""
-    config = Gemma4VisionConfig
+    config: Gemma4VisionConfig

Suppressing this rule

Add a # trf-ignore: TRF055 comment on the flagged line or the line directly above it. See Suppressing rules for whole-file directives and when a suppression is the wrong answer.