TRF051
Modeling code must not branch on _attn_implementation; the attention interface owns dispatch.
| Default | Enabled |
| Scope | All models |
| Source | mlinter/trf051.py |
| Show in terminal | mlinter --rule TRF051 |
What it does
Checks modeling_*.py and modular_*.py for comparisons against a _attn_implementation attribute.
Why is this bad?
Backend dispatch belongs to ALL_ATTENTION_FUNCTIONS.get_interface, and backend-conditional tensor munging (padding, reshaping) belongs in the shared wrappers under integrations/. Inline branching keeps the model body kernel-aware and breaks when new backends register.
Example
-if self.config._attn_implementation == "flash_attention_2":
- attn_output = flash_path(query_states, key_states, value_states)
-else:
- attn_output = eager_path(query_states, key_states, value_states)
+attention_interface = ALL_ATTENTION_FUNCTIONS.get_interface(self.config._attn_implementation, eager_attention_forward)
+attn_output, attn_weights = attention_interface(self, query_states, key_states, value_states, ...)
Suppressing this rule
Add a # trf-ignore: TRF051 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
9 models are exempt from TRF051 in mlinter/rules.toml, because they predate the convention and cannot be changed without breaking backward compatibility.
Show the 9 allowlisted models
axk2decision_transformerdeepseek_v32diffusion_gemmafalconglm_moe_dsagpt2minimax_m3_vlqwen2_5_omni