Ollama is not a wrapper around llama.cpp anymore, at least for multimodal models (not sure about others). They have their own engine: https://ollama.com/blog/multimodal-models