TRF058
Buffers must be declared as nn.Buffer attributes, not registered with register_buffer().
| Default | Enabled |
| Scope | All models |
| Source | mlinter/trf058.py |
| Show in terminal | mlinter --rule TRF058 |
What it does
In modeling_*.py and modular_*.py, flags register_buffer("<name>", ...) calls whose buffer name is a string literal, on any receiver (self, or another module such as layer.mamba). A computed name – a variable or f-string, e.g. one buffer per layer inside a loop – has no attribute-assignment equivalent and is exempt.
Why is this bad?
Since torch>=2.5 nn.Buffer registers a buffer through plain attribute assignment, like nn.Parameter. A buffer created by a method call only exists as a side effect of running __init__, so a modular file that wants to tweak one has to redefine the whole __init__. Assigned as an attribute, it can be inherited and overridden on its own.
Example
- self.register_buffer("inv_freq", inv_freq, persistent=False)
- self.register_buffer(
- "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
- )
+ self.inv_freq = nn.Buffer(inv_freq, persistent=False)
+ self.position_ids = nn.Buffer(torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False)
Suppressing this rule
Add a # trf-ignore: TRF058 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
2 models are exempt from TRF058 in mlinter/rules.toml, because they predate the convention and cannot be changed without breaking backward compatibility.
Show the 2 allowlisted models
falcon_h1pp_doclayout_v2