TRF057
Public model, config, output and processor classes and their public methods must be decorated with @auto_docstring.
| Default | Enabled |
| Scope | Models added on or after 2026-06-20 |
| Source | mlinter/trf057.py |
| Show in terminal | mlinter --rule TRF057 |
What it does
Checks @auto_docstring on the classes that need it: public PreTrainedModel subclasses (<Model>PreTrainedModel, <Model>Model, <Model>For<Task>, backbones), PreTrainedConfig subclasses, ModelOutput subclasses, image processors and ProcessorMixin subclasses, and on their public methods: forward, get_image_features, get_video_features, get_audio_features, get_text_features, preprocess and __call__. A class or method in a modular_*.py file is checked against the files generated from it.
Why is this bad?
Without the decorator, a class ships with no intro and no parameter documentation, and a method with no argument documentation, no Returns section and no usage example, so the standard descriptions in auto_docstring.py have to be hand-written per model instead.
Example
+@auto_docstring
@dataclass
class AcmeModelOutputWithPast(ModelOutput):
logits: torch.FloatTensor | None = None
+@auto_docstring
class AcmeForConditionalGeneration(AcmePreTrainedModel):
+ @auto_docstring
def forward(self, input_ids, pixel_values=None, **kwargs):
...
Suppressing this rule
Add a # trf-ignore: TRF057 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.
Allowlisted models
6 models are exempt from TRF057 in mlinter/rules.toml, because they predate the convention and cannot be changed without breaking backward compatibility.
Show the 6 allowlisted models
kosmos2timm_backbonetimm_wrappervitpose_backbonewav2vec2_with_lmx_clip