Support multimodal models such as LLaVA for image input

[cebtenzzre]

Feature request

We can make use of the upstream work at ggerganov/llama.cpp#3436 to support image input to LLMs.

@AndriyMulyar What was the name of the model that you wanted to consider as an alternative to LLaVA?

Motivation

Real-time image recognition on resource-constrained hardware would be very useful in applications such as robotics. This feature would open the door to broader use cases for GPT4All than simple text completion.

Your contribution

I may submit a pull request implementing this functionality.

[AndriyMulyar]

Fuyu 8b is interesting because its decoder only.

I think LLaVA style is a fine choice though for an initial multimodal implementation

[manyoso]

This will require extensive changes to the GUI as well. It has been agreed that the GUI changes will come first to provide a UI for the current multimodel upstream.

[PedzacyKapec]

+1