examples/server: Return currently-loaded model

[vtmmm]

I'm not making a pull request because I'm not sure if this is something that should be default behavior, or if it should be implemented as an option.

I'm using the server example, but would like it to return the currently-loaded model with each transcription. This is in case the /load endpoint is used, and the model has changed since server was run.

Here are the changes I made to implement this for the default JSON type response. I know the Verbose JSON is intended to match OpenAI's, so I don't know whether the community would like it added there or not.

In server.cpp:

• At line 914, update JSON response by adding the "model" field.

            json jres = json{
                {"text", results},
                {"model", default_params.model.c_str()}
            };

• Add the following line before line 958 so that the params data structure is updated with the new model after it is verified to exist:

default_params.model = model.c_str();

Thoughts/feedback?

Updated to use default_params variable instead of params

[felrock]

Perhaps it would make sense to add an endpoint to check which is the currently loaded, i.e /currently_loaded_model. Which returns a json object containing the model name and parameters. I think its a bit of an overhead to have with each inference request.

[vtmmm]

Yes, I think an endpoint like that would be a good idea. Maybe a /model endpoint, that behaved like /load with a PUT method or returned the model/parameters with a GET method.

If someone has the use case where they want to verify the model for every transcription, then making two API calls per transcription also adds overhead (since other parameters can be specified with the request). But that might be a rare use case, and plus it's an easy patch to add it.

I also don't know how long the transcriptions are that most people are running per request, and therefore how much relative overhead there is by including the filepath to the loaded model.