Fine-tune service designing

[Aisuko]

Currently, fine-tuning feature only for specific model(small language model). As my research, HuggingFace Trainer isn't a good way we embed it as the training function. So, here is what I'm thinking:
We should use PyTorch to do the fine-tuning process rather than HuggingFace Trainer(inlcuded the trainer from trl). It will take a long time for us to implement, but the APIs behaviours are more controllable.