TRF024

Layer dimensions must come from the config, not from an integer literal in the modeling file.

Default	Enabled
Scope	Models added on or after 2026-06-20
Source	mlinter/trf024.py
Show in terminal	mlinter --rule TRF024

What it does

In modeling_*.py and modular_*.py, flags an integer literal greater than 8 in a dimension argument of a torch.nn constructor (Linear, Embedding, LayerNorm, RMSNorm, GroupNorm, BatchNorm*, InstanceNorm*, Conv*d, ConvTranspose*d, Bilinear, MultiheadAttention), positional or by keyword (in_features, out_features, in_channels, out_channels, num_embeddings, embedding_dim, embed_dim, normalized_shape, num_channels, hidden_size). Operator-shape arguments (kernel_size, stride, padding, num_groups) are ignored; literals up to 8 keep scalar heads, binary classifiers and RGB channel counts clean. Models added before cutoff_date are exempt.

Why is this bad?

A hardcoded width pins the module to one checkpoint size: the same architecture at another scale loads with a shape mismatch, and from_pretrained cannot say which value is wrong because no config field points at it. It also splits the source of truth, so editing the config no longer changes the model that gets built.

Example

 class AcmeAtomEmbedding(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.proj = nn.Linear(768, 3072, bias=False)
-        self.norm = nn.LayerNorm(3072)
+        self.proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
+        self.norm = nn.LayerNorm(config.intermediate_size)

Suppressing this rule

Add a # trf-ignore: TRF024 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.