server: remove default "gpt-3.5-turbo" model name

[ngxson]

Fix #17666

Important

This PR changes the way how model name is handle (see below).
In addition to that, the "model" field from input request no longer be reflected back to the response. If you need to use a custom model name, set it via --model-alias argument instead

Remove the default gpt-3.5-turbo model name as was useful in the past, but no longer be the case.

The model name is now decided based on this priority:

• If model_alias is set, use it
• Otherwise, use registry format <user>/<model>:<tag> if available (for cached models)
• Fallback to file name if none is found

After the change, web UI display the model name correctly:

llama-server -m ../models/DeepSeek-R1-Distill-Llama-8B-Q4_K_M.gguf

[pwilkin]

LGTM, should be a generally safer approach anyway.

[pwilkin]

You might want to also check the stats under the assistant response, I think that's where it bugged out for me (if you have 'show statistics' and 'show model name' checked in settings).

[ngxson]

Hmm yeah, it reflects back the input model name from the request. I think it's also time to remove this behavior as it doesn't make much sense.

[ngxson]

@pwilkin could you confirm if the latest commit resolves the problem on your side? (maybe there are cases that I haven't tested)

[pwilkin]

Okay, pulling and will check.

[pwilkin]

@ngxson can confirm, works good now.

[ngxson]

Merging this once most workflows passed on the mirrored PR: ngxson#46