TRF046

forward must not write module attributes; modules are stateless in forward.

Default	Enabled
Scope	All models
Source	mlinter/trf046.py
Show in terminal	mlinter --rule TRF046

What it does

Checks forward methods in modeling_*.py and modular_*.py for assignments to self attributes.

Why is this bad?

Hidden state written during forward breaks batching, torch.compile, and reasoning about the module. Carried state is passed explicitly (cache objects, the generate loop); values that depend only on config or static shapes belong in __init__.

Example

 def forward(self, hidden_states):
-    self.sequence_length = hidden_states.shape[1]
-    embeddings = self.compute_embeddings(self.sequence_length)
+    embeddings = self.compute_embeddings(hidden_states.shape[1])

Suppressing this rule

Add a # trf-ignore: TRF046 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

13 models are exempt from TRF046 in mlinter/rules.toml, because they predate the convention and cannot be changed without breaking backward compatibility.

Show the 13 allowlisted models

• clvp
• codegen
• ctrl
• fastspeech2_conformer
• nllb_moe
• qwen2_5_omni
• qwen3_omni_moe
• recurrent_gemma
• seamless_m4t
• switch_transformers
• trocr
• udop
• wav2vec2_conformer