Support multimodals models with vLLM

[mudler]

Is your feature request related to a problem? Please describe.
Many models are now becoming multi-model, that is they can accept images, videos or audio during inference. The llama.cpp project is currently providing multimodal support and we do as well by using it, however there are models which aren't supported yet (for instance #3535 and #3669, see also ggml-org/llama.cpp#9455 )

Describe the solution you'd like
LocalAI to support vLLM multimodal capabilities

Describe alternatives you've considered

Additional context
See #3535 and #3669, tangentially related to: #2318 #3602

https://github.com/vllm-project/vllm/blob/main/examples/offline_inference_pixtral.py

https://github.com/vllm-project/vllm/blob/main/examples/offline_inference_vision_language_multi_image.py

[3unnycheung]

agree

[SuperPat45]

I am very interested in the support of vision models in localAI particularly Llama-3.2-11B-Vision and Pixtral-12b

[mudler]

With #3729 should cover most of the models and add also video understanding. Model configuration files needs to be specify placeholders used by models for image/video tags in the text prompt, going to experiment with this once in master and update the model gallery with few examples.

[AlexM4H]

Great news. However, I miss two Docker images
master-cublas-cuda12-ffmpeg and master-aio-gpu-nvidia-cuda-12