transformer_lens.model_bridge.supported_architectures.llama4_multimodal module

Llama 4 multimodal architecture adapter.

Meta’s Llama 4 composite (Llama4ForConditionalGeneration): a Llama4 vision transformer at vision_model and a projector feeding the full Llama4ForCausalLM at language_model (text stack at language_model.model.* with language_model.lm_head). The vision side is delegated opaquely; the text mapping is the Llama4 text adapter’s, re-prefixed.

class transformer_lens.model_bridge.supported_architectures.llama4_multimodal.Llama4MultimodalArchitectureAdapter(cfg: Any)

Bases: Llama4ArchitectureAdapter

Architecture adapter for Llama4ForConditionalGeneration models.

__init__(cfg: Any) → None

Initialize the Llama 4 multimodal architecture adapter.