TRF041
A config-gated branch must carry a # CODEPATH: note saying which checkpoints diverge.
| Default | Enabled |
| Scope | Models added on or after 2026-06-20 |
| Source | mlinter/trf041.py |
| Show in terminal | mlinter --rule TRF041 |
What it does
In modeling_*.py and modular_*.py, flags every if/elif statement and every conditional expression whose condition reads a config.* or self.config.* attribute and that does not carry a # CODEPATH: comment. The comment is accepted on the branch line itself or anywhere in the contiguous comment block directly above it, so it can head a multi-line explanation. Deliberately broad: any config attribute in the condition counts, not only boolean feature flags, because a branch on a numeric or optional config field forks the graph exactly as much as a branch on a flag does. One shape is exempt outright, by structure rather than by name: X if X is not None else fallback, where the field under test is itself one of the two results. That yields the field when set and a default when not, which is getattr(config, x, default) spelled long, so it cannot fork the graph and there is no path to name. Merely mentioning None does not qualify — config.vision_config is not None selects a whole extra tower and still owes a note. A field that gates no checkpoint divergence at all, such as problem_type selecting a loss or hidden_act looking up an activation, can be exempted for a whole file with a module-level # trf-ignore: TRF041 config.problem_type, config.hidden_act directive at column 0, naming the fields comma- or space-separated; self.config.x, config.x and x all name the same field. The directive has to name at least one field, so a bare # trf-ignore: TRF041 keeps its per-line meaning instead of muting the file. Exemption is per field, not per branch: a condition reading several config fields is skipped only when every one of them is exempt, so if config.problem_type and config.use_cache still has to explain itself when only problem_type is named.
Why is this bad?
Every config-gated branch is a second architecture living in the same file, and the reader cannot tell from the code whether both halves are reachable. That is why reviewers ask “is this ever used?”, “are they all needed?” and “why are there so many cases?” on almost every new model, and why dead experimental branches survive for releases. This rule does not forbid the branch; it borrows Rust’s // SAFETY: discipline and makes the author write down the justification next to it. A branch nobody can name a checkpoint for is a branch to delete, and the note makes that obvious at review time instead of three rounds later.
Example
+ # CODEPATH: ESMC-6B ships pre-normalised embeddings, the 300M/600M checkpoints do not.
if config.use_embedding_norm:
hidden_states = self.embedding_norm(hidden_states)
- if config.msa_encoder_enabled:
- hidden_states = self.msa_encoder(hidden_states)
+ # no released checkpoint sets msa_encoder_enabled -> branch removed
Suppressing this rule
Add a # trf-ignore: TRF041 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
11 models are exempt from TRF041 in mlinter/rules.toml, because they predate the convention and cannot be changed without breaking backward compatibility.
Show the 11 allowlisted models
dinov3_vitinklingkimi_k25kosmos2mimo_v2_flashopenaitimm_backbonetimm_wrappertipsv2vitpose_backbonex_clip