TRF052
No *_ATTENTION_CLASSES dispatch dicts; attention backends route through the attention interface.
| Default | Enabled |
| Scope | All models |
| Source | mlinter/trf052.py |
| Show in terminal | mlinter --rule TRF052 |
What it does
Checks modeling_*.py and modular_*.py for module-level assignments to names ending in _ATTENTION_CLASSES.
Why is this bad?
Per-backend attention classes selected from a dict are the pre-interface idiom: three near-identical classes drift apart, and hub attention kernels registered into ALL_ATTENTION_FUNCTIONS never reach them. One attention class dispatching through the interface replaces the dict; do not propagate it from a legacy parent.
Example
-ACME_ATTENTION_CLASSES = {
- "eager": AcmeAttention,
- "flash_attention_2": AcmeFlashAttention2,
- "sdpa": AcmeSdpaAttention,
-}
+class AcmeAttention(nn.Module):
+ def forward(self, ...):
+ attention_interface = ALL_ATTENTION_FUNCTIONS.get_interface(self.config._attn_implementation, eager_attention_forward)
Suppressing this rule
Add a # trf-ignore: TRF052 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
8 models are exempt from TRF052 in mlinter/rules.toml, because they predate the convention and cannot be changed without breaking backward compatibility.
Show the 8 allowlisted models
barkdata2vecfalcongitgpt_neogptjsamsuperglue