TRF018

_init_weights overrides should call super()._init_weights(module), except modular-file sentinels.

Default	Enabled
Scope	All models
Source	mlinter/trf018.py
Show in terminal	mlinter --rule TRF018

What it does

Checks that every PreTrainedModel subclass overriding _init_weights(self, module, ...) chains up via super()._init_weights(...). Modular files also accept the sentinels PreTrainedModel._init_weights(self, module), PreTrainedModel._init_weights(module) and raise AttributeError(...). For a deliberate full override, suppress with # trf-ignore: TRF018 above the method.

Why is this bad?

The base _init_weights covers the standard module types (Linear, Embedding, LayerNorm, RotaryEmbedding, …). Skipping it leaves every submodule the override misses uninitialized – which passes tests and surfaces much later as a subtle weight-init bug (cf. https://github.com/huggingface/transformers/pull/45597).

Example

 from ... import initialization as init

 def _init_weights(self, module):
+    super()._init_weights(module)
     if isinstance(module, AcmeCustomLayer):
-        module.gate.data.zero_()
+        init.zeros_(module.gate)

Suppressing this rule

Add a # trf-ignore: TRF018 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.