All chat templates seem broken?

[Kwisss]

As the title says, every chat template seems to be broken in llama-server?.
It VERY possible that the issue is between the chair en the keyboard (me)
but I do not know what to do from here? Its not some super resent change since I tried building slightly older versions.

I asked the SOTA models (gemini) ofcourse but I couldn't get it fixed?

I tried different sizes of Qwen, from bartowkki and unsloth and some random models from bartowksi.
but after 2 or 3 messages the models start responding to themself, or mixing up who said what?

I'm adding " --jinja" but also tried without, I even tried adding "--reasoning-format deepseek" for qwen, and it is a bit different but still responds on its own messages?

./llama-server.exe --model Qwen3-30B-A3B-Thinking-2507-UD-Q8_K_XL.gguf --jinja -ngl 99 --threads -1 --ctx-size 131072 --temp 0.7 --min-p 0.0 --top-p 0.80 --top-k 20 --presence-penalty 1.0 --cache-type-k q8_0 --cache-type-v q8_0 --flash-attn on -ot ".ffn_(up|down)_exps.=CPU"

[JO-C-Debug]

I've encountered this situation too. Adding a template sometimes causes it to ramble incoherently and even display strange symbols.Especially when discussing restricted topics

[openmarmot]

not sure if this is your exact issue but I found on newer versions of llama.cpp i had to add "--chat-template chatml" to a mistral based model that had previously just worked. without this it would ignore the prompt and just talk about random weird stuff.

[ali0une]

Hi there.

i have a similar issue when adding --chat-template gemma the server answers "unable to tokenize prompt" and crashes.