Eval bug: Regression: --reasoning-budget N is ignored

[shikaku2]

Name and Version

working build (might not be where it regression was introduced, don't have time to test today, can update later):
llama-server --version
version: 8661 (b7ad48e)
built with GNU 15.2.1 for Linux x86_64

nonworking build:
llama-server --version
version: 8668 (b7ad48e)
built with GNU 15.2.1 for Linux x86_64

Operating systems

Linux

GGML backends

Vulkan

Hardware

n/a, general bug

Models

gemma-4-26b-a4b-it tested, might affect more

Problem description & steps to reproduce

in nonworking build(s), --reasoning-budget N is ignored entirely
llama-server -m gemma-4-26B-A4B-it-UD-IQ4_XS.gguf --reasoning-budget 100 --reasoning-budget-message "STOP"
[STOP never triggers]

First Bad Commit

No response

Relevant log output

[no crashes or abnormal logs, it's just a command line argument being ignored]

[aldehir]

Only affects Gemma4 because it wasn't added in. That said, you should first try to adjust reasoning in accordance to their prompting guide: https://ai.google.dev/gemma/docs/core/prompt-formatting-gemma4#adaptive-thought-efficiency