Allow for changing shared.settings during model load via API

[brthor]

Checklist:

• I have read the Contributing guidelines.

Reasoning

Sometimes, when you load a certain lora, you need to be able to change the instruction_template on the fly.

Rather than focus on that specific use case, allow for updating the shared.settings as a whole at the time of model load via the api.

[oobabooga]

The current API will be replaced with an OpenAI-compatible one, and handling instruction templates is an open problem there. In particular, making /v1/chat/completions work properly when different models are loaded. Can you take a look at #4430? PRs to the openai-api branch are welcome.

[oobabooga]

Seems to be working for the OpenAI API:

  curl -k http://127.0.0.1:5000/v1/internal/model/load \
    -H "Content-Type: application/json" \
    -d '{
      "model_name": "facebook_galactica-125m",
      "args": {
        "load_in_4bit": true
      },
      "settings": {
        "instruction_template": "Alpaca"
      }
    }'