server: Retrieve prompt template in /props

[bviksoe]

This PR adds the following:

• Expose the actual Jinja2 prompt template of the model in the /props endpoint.
• Change log-level from Error to Warning for warning about template mismatch.

The front-end stands a better chance of actually executing the Jinja template format correctly. Server is currently just guessing it.

Ideally this should have been inside a JSON block that expose the same key/value pairs as listed during startup in llm_load_print_meta function, allowing front-end to read a plethora of model properties.

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

[ngxson]

I'm not very confident about this change. Even when frontend can get the jinja template, how it will be used? I suppose it requires a jinja parser on frontend?

Another way I can think of is that backend can format a chat template with some pre-defined placeholders, like {{user_message}}, and the JS code can do simple string manipulation to get the correct template.

[bviksoe]

Even when frontend can get the jinja template, how it will be used? I suppose it requires a jinja parser on frontend?

Yes. Adding a real Jinja parser in c++ has already been discussed I think. It's not likely to be added. But lots of Jinja exist for Python and JavaScript.

The /chat/completions endpoint will surely help standard usage of server, and it does its assumptions about the template content. But for those on the /completion endpoint will want to have access to all model properties, including the template - so they can ensure any model prompt formats correctly.