TRF015

Models with non-empty _tied_weights_keys must have tie_word_embeddings in their Config.

Default	Enabled
Scope	All models
Source	mlinter/trf015.py
Show in terminal	mlinter --rule TRF015

What it does

When a PreTrainedModel subclass sets a non-empty _tied_weights_keys, checks the companion configuration file for a tie_word_embeddings field.

Why is this bad?

Without it users cannot control weight tying: the model ties unconditionally, breaking serialization round-trips and fine-tuning with untied heads.

Example

 # configuration_foo.py
 @strict(accept_kwargs=True)
 class FooConfig(PreTrainedConfig):
     hidden_size: int = 768
+    tie_word_embeddings: bool = True

Suppressing this rule

Add a # trf-ignore: TRF015 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.