Add modal training capability for each model #33
Description
Script up ways of training, deploying and inferencing models on Modal.
Should have:
- one key model backend. We'll choose the easiest one to use that works well (something along the lines of a llama)
- lora saving capability
- lora volume logging capability
- load correct lora and inference (ensure this doesn't add too much latency)
Nice to have:
- github actions to automate training when data sources train
- monitoring during training to understand performance (maybe weights and biases?)
- understanding of model performance against third party APIs