Coverage for transformer_lens/model_bridge/supported_architectures/marian.py: 100%

1"""Marian architecture adapter. 

2 

3Helsinki-NLP's opus-mt family (``MarianMTModel``): post-norm Bart layout with 

4no layernorm_embedding and no per-stack final norms, deterministic sinusoidal 

5positions, and the sqrt(d_model) embedding scale applied in the stack forward 

6(embed hooks observe unscaled output). A trained ``final_logits_bias`` is 

7added after lm_head inside HF's forward. 

8""" 

9 

10from transformer_lens.model_bridge.supported_architectures.bart import ( 

11 BartFamilyArchitectureAdapter, 

12) 

13 

14 

15class MarianArchitectureAdapter(BartFamilyArchitectureAdapter): 

16 """Architecture adapter for MarianMTModel models (Helsinki-NLP opus-mt family)."""