transformer_lens.model_bridge.supported_architectures.llava_onevision module¶

LLava-OneVision architecture adapter.

Same module hierarchy as base LLava; SigLIP encoder and Qwen2 backbone are handled dynamically by the base adapter and HuggingFace’s forward().

class transformer_lens.model_bridge.supported_architectures.llava_onevision.LlavaOnevisionArchitectureAdapter(cfg: Any)¶

Architecture adapter for LLaVA-OneVision models.

prepare_model(hf_model: Any) → None¶

Fix weight tying when text_config and top-level config disagree.

Some checkpoints have tie_word_embeddings=True in text_config but False at the top level, leaving lm_head randomly initialized.