Multiple model aliases (vLLM like)

[ELigoP]

This is very useful if you want to serve to models with different params (prompt, or temperate etc.) in some frontend, e.g. OpenWebUI. I use Qwen3 models, and I want to serve same model with /no_think prompt and without this (thinking on). And even recommended temperature and top-p is different for thinking and non-thinking mode.

Currently if you enter multiple --alias X --alias Y values, last Y value is used.
vLLM allows --served-model-name X Y, then both X and Y are returned from /v1/models request, both can be called.