Add Finetune Support

[alex4321]

In the late versions of llama.cpp there are ways to fine-tune the model using LoRA adapters: ggerganov/llama.cpp#2632

However, either I missed it or are there no functions introduced in this PR imported into this project?

Describe alternatives you've considered
Sure it's always an option to use a CLI tool, but I guess Python binding can be useful.

[abetlen]

@alex4321 not yet but I am planning to work on this. The first step is actually exposing the training API in llama.cpp as it's currently only available as an example which requires working with the GGML API directly.

[satyaloka93]

Hi, will there be a way to switch lora adapters, like PEFT enable/disable feature?

[nai-kon]

Hi, will there be a way to switch lora adapters, like PEFT enable/disable feature?

I want to know it too. It's like load_adapter/set_adapter of transformers library.