WebUI: shouldn't override server side sampling settings by default

[notgood]

SvelteKit WebUI always overrides server side configured model sampling settings like temperature, top_p, top_k, min_p and many others.

Each model typically recommends it's own temperature and sampling settings, so for example running gpt-oss as suggested in @ggerganov guide #15396
llama-server --temp 1.0 --top-p 1.0 --top-k 0 --min-p 0.001 --model gpt-oss-20b-mxfp4.gguf

Yet with WebUI settings reset to default, following parameters are sent to server:

srv log_server_r: request: {"messages":[{"role":"user","content":"hi"}],"stream":true,"reasoning_format":"auto","temperature":0.8,"max_tokens":-1,"dynatemp_range":0,"dynatemp_exponent":1,"top_k":40,"top_p":0.95,"min_p":0.05,"xtc_probability":0,"xtc_threshold":0.1,"typ_p":1,"repeat_last_n":64,"repeat_penalty":1,"presence_penalty":0,"frequency_penalty":0,"dry_multiplier":0,"dry_base":1.75,"dry_allowed_length":2,"dry_penalty_last_n":-1,"samplers":["top_k","typ_p","top_p","min_p","temperature"],"timings_per_token":true}

After editing WebUI settings and leaving most of the relevant input fields empty, zeroes are sent to server, which probably is a bug.

srv log_server_r: request: {"messages":[{"role":"user","content":"hi"}],"stream":true,"reasoning_format":"auto","temperature":0,"max_tokens":-1,"dynatemp_range":0,"dynatemp_exponent":0,"top_k":0,"top_p":0,"min_p":0,"xtc_probability":0,"xtc_threshold":0,"typ_p":0,"repeat_last_n":0,"repeat_penalty":0,"presence_penalty":0,"frequency_penalty":0,"dry_multiplier":0,"dry_base":0,"dry_allowed_length":0,"dry_penalty_last_n":0,"timings_per_token":true}

Expected default behaviour would be empty relevant fields in WebUI settings and only minimal required params sent to server, thus respecting server-side model configuration.

{"messages":[{"role":"user","content":"hi"}],"stream":true,"timings_per_token":true}

Overriding model settings WebUI side should be concious opt-in decision by user, sending a copy of hardcoded values by default is rather confusing.

[openingnow]

Another option is to include a button that lets users choose whether to skip parameters. #13277 (comment) (from prior webui)

[ServeurpersoCom]

OK, now that we can explicitly send 0 when we want to, we also need the ability to send nothing at all if the corresponding UI field is left empty. That's the true OpenAI-Compat behavior: either you override, or you leave it blank.

Currently, if I clear a sampling parameter; for example, the temperature; the outgoing request looks like this:

..."reasoning_format":"auto","temperature":0,"max_tokens":-1,"dynatemp_range":0,"dynatemp_exponent":1,"top_k":40,"top_p":0.95,"min_p":0.05,"xtc_probability":0...

but it should be:

..."reasoning_format":"auto","max_tokens":-1,"dynatemp_range":0,"dynatemp_exponent":1,"top_k":40,"top_p":0.95,"min_p":0.05,"xtc_probability":0...

[ServeurpersoCom]

I'm currently testing this commit:
affee99
I'll open a PR once I’ve validated it thoroughly.

[notgood]

@ServeurpersoCom thank you, it's a big step in the right direction!

I've tested these changes, the only problem is a handling of "max_tokens": empty field in GUI is still being sent to server as -1 default.

log_server_r: request: {"messages":[{"role":"user","content":"hi"}],"stream":true,"reasoning_format":"auto","max_tokens":-1,"timings_per_token":true}

Rest of the empty sampling-related settings are being ignored, perfect!

[ServeurpersoCom]

I've tested these changes, the only problem is a handling of "max_tokens": empty field in GUI is still being sent to server as -1 default.

log_server_r: request: {"messages":[{"role":"user","content":"hi"}],"stream":true,"reasoning_format":"auto","max_tokens":-1,"timings_per_token":true}

Rest of the empty sampling-related settings are being ignored, perfect!

Oh, I missed that one! Thanks for catching it: I'll take a look 👀

[ServeurpersoCom]

#16489

Testing

max_tokens Empty (let server decide)

{"messages":[{"role":"user","content":"Salut"}],"stream":true,"model":"Llama-3_3-Nemotron-Super-49B-v1_5","reasoning_format":"auto","temperature":0.8,"dynatemp_range":0,"dynatemp_exponent":1,"top_k":40,"top_p":0.95,"min_p":0.05,"xtc_probability":0,"xtc_threshold":0.1,"typ_p":1,"repeat_last_n":64,"repeat_penalty":1,"presence_penalty":0,"frequency_penalty":0,"dry_multiplier":0,"dry_base":1.75,"dry_allowed_length":2,"dry_penalty_last_n":-1,"samplers":["top_k","typ_p","top_p","min_p","temperature"],"timings_per_token":true}

max_tokens WebUI Default :

{"messages":[{"role":"user","content":"Salut"}],"stream":true,"model":"Llama-3_3-Nemotron-Super-49B-v1_5","reasoning_format":"auto","temperature":0.8,"max_tokens":-1,"dynatemp_range":0,"dynatemp_exponent":1,"top_k":40,"top_p":0.95,"min_p":0.05,"xtc_probability":0,"xtc_threshold":0.1,"typ_p":1,"repeat_last_n":64,"repeat_penalty":1,"presence_penalty":0,"frequency_penalty":0,"dry_multiplier":0,"dry_base":1.75,"dry_allowed_length":2,"dry_penalty_last_n":-1,"samplers":["top_k","typ_p","top_p","min_p","temperature"],"timings_per_token":true}

[notgood]

Confirming, with a latest build and max_tokens empty in webui settings, nothing being sent to server

log_server_r: request: {"messages":[{"role":"user","content":"hi"}],"stream":true,"reasoning_format":"auto","timings_per_token":true}