[Chatllama] Support Inference for trained models.

[PierpaoloSorbellini]

Description

Currently to perform inference of the models generated the user needs to interact with the model generated writing a small python script accordingly to how the model is saved by library, by loading the resulting checkpoint or model saved after training.

Moreover a lot of optimization can be integrated to speed-up the inference such as:

• CPU Offloading.
• llama.ccp implementation
• accelerate / deepspeed distributed inference.

TODO

• Implement Inference Class to make inference very easy and even possible from CLI.
• Implement Inference with the optimisations available from deepspeed
• Implement inference with the optimisations available from accelerate
• Implement fast lama inference with known library llama.ccp implementation

[shrinath-suresh]

@PierpaoloSorbellini The inference section is tagged with WIP. Do we have any basic inference code available in chatllama to load actor_rl model and run few queries ?