llama.cpp: Do server launch parameters or WebUI settings control sampling during inference?

[sidran]

I'm running llama.cpp server on Windows with custom sampling parameters (e.g., --temp, --top-k, etc.).
After launching, the WebUI displays different values for these parameters and they can be adjusted there.
During inference, which settings are actually used: the ones passed at server startup, or the ones shown/changed in the WebUI?
Trying to understand which layer has final control.
Thanks

[ggerganov]

The WebUI overrides the default values that you provide to llama-server.

[sidran]

Thank you.

Can you just take a glimpse at this thread and my suggestions presented there? WebUI improvements
Can something be done about it?