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

1"""Youtu architecture adapter. 

2 

3Tencent's Youtu-LLM (``YoutuForCausalLM``, native in transformers): the 

4only laptop-scale densely-MLP'd Multi-head Latent Attention checkpoint — 

5DeepSeek-V2's MLA (q LoRA + compressed KV with decoupled rope) with every 

6layer dense. Module names match DeepSeek-V2 exactly and the MLA bridge 

7already handles both q-projection variants, so this is a pure subclass: 

8the MoE router/shared-expert submodules are optional and simply never 

9bind on the dense MLPs. 

10""" 

11 

12from transformer_lens.model_bridge.supported_architectures.deepseek_v2 import ( 

13 DeepSeekV2ArchitectureAdapter, 

14) 

15 

16 

17class YoutuArchitectureAdapter(DeepSeekV2ArchitectureAdapter): 

18 """Architecture adapter for YoutuForCausalLM models.""" 

19 

20 def _build_mlp_bridge(self): 

21 """Every layer is a plain gated MLP — map it so weight processing 

22 sees mlp.in/gate/out (the parent's MoE wrapper leaves them opaque).""" 

23 return self._gated_mlp()