Is it possible to use llama.cpp engine to generate with LLM from embeddings and not from tokens?

[Ouna-the-Dataweaver]

So, basically, for multimodal models you quite often need to generate straight from embeddings. You take in prompt = tokens, apply embedding layer to them to get your tensor for text, and insert inside embeddings tensors from other modality encoders. In pytorch you can directly pass embeddings into LLM like that:

llama_model.generate(
            inputs_embeds=embeds,...)

Is there any way to do something similar in llama.cpp?