transformer_lens.model_bridge.supported_architectures.youtu module¶
Youtu architecture adapter.
Tencent’s Youtu-LLM (YoutuForCausalLM, native in transformers): the
only laptop-scale densely-MLP’d Multi-head Latent Attention checkpoint —
DeepSeek-V2’s MLA (q LoRA + compressed KV with decoupled rope) with every
layer dense. Module names match DeepSeek-V2 exactly and the MLA bridge
already handles both q-projection variants, so this is a pure subclass:
the MoE router/shared-expert submodules are optional and simply never
bind on the dense MLPs.
- class transformer_lens.model_bridge.supported_architectures.youtu.YoutuArchitectureAdapter(cfg: Any)¶
Bases:
DeepSeekV2ArchitectureAdapterArchitecture adapter for YoutuForCausalLM models.