Support Multimodal model

[Aisuko]

• Kirin project should support GPU first
• Quantized multimodel LLM
• UI support load image
• API support tokenised image

Same to before, we will make choice on https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/tree/main. And quantized the model to Q4_x. Although there are articles state that quantization would bad for multimodal model accuracy. However, we are at the experiment stag, it should be ok.

And after we migrate the base model to multimodal, the basic requirements of the local environment should be also upgraded.

Currently:

• PHI3-vision gguf conversion ggml-org/llama.cpp#7705
• https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/discussions/40
• FR: Phi-3-vision-128k-instruct implementation ggml-org/llama.cpp#7444
• Support multimodal LLMs? huggingface/candle#1947

It's absolutely for us beginning our multimodal journey.

[Aisuko]

https://github.com/SkywardAI/kirin/milestone/25