More Gemma4 fixes in the past 24 hours

Reasoning budget fix (merged): https://github.com/ggml-org/llama.cpp/pull/21697

New chat templates from Google to fix tool calling:

31B: https://huggingface.co/google/gemma-4-31B-it/blob/main/chat_template.jinja

27B: https://huggingface.co/google/gemma-4-26B-A4B-it/blob/main/chat_template.jinja

E4B: https://huggingface.co/google/gemma-4-E4B-it/blob/main/chat_template.jinja

E2B: https://huggingface.co/google/gemma-4-E2B-it/blob/main/chat_template.jinja

Please correct me if Im wrong, but you should use these new templates unless you redownload a new GGUF, that has been updated in the past 24 hours with the new template.

You can use specific templates in llama.cpp by the command argument:

--chat-template-file /models/gemma4/gemma4_chat_template_26B.jinja 

My current llama-swap/llama.cpp config 26B example (testing on 16GB VRAM , so context window is limited):

"Gemma4-26B-IQ4_XS": ttl: 300 # Automatically unloads after 5 mins of inactivity cmd: > /usr/local/bin/llama-server --port ${PORT} --host 127.0.0.1 --model /models/gemma4/gemma-4-26B-A4B-it-UD-IQ4_XS.gguf --mmproj /models/gemma4/gemma-4-26B-A4B-it.mmproj-q8_0.gguf --chat-template-file /models/gemma4/gemma4_chat_template_26B_09APR2026.jinja --cache-type-k q8_0 --cache-type-v q8_0 --n-gpu-layers 99 --parallel 1 --batch-size 2048 --ubatch-size 512 --ctx-size 16384 --image-min-tokens 300 --image-max-tokens 512 --flash-attn on --jinja --cache-ram 2048 -ctxcp 2 filters: stripParams: "temperature, top_p, top_k, min_p, presence_penalty, repeat_penalty" setParamsByID: "${MODEL_ID}:thinking": chat_template_kwargs: enable_thinking: true reasoning_budget: 4096 temperature: 1.0 top_p: 0.95 top_k: 64 min_p: 0.0 presence_penalty: 0.0 repeat_penalty: 1.0 "${MODEL_ID}:thinking-coding": chat_template_kwargs: enable_thinking: true reasoning_budget: 4096 temperature: 1.5 top_p: 0.95 top_k: 65 min_p: 0.0 presence_penalty: 0.0 repeat_penalty: 1.0 "${MODEL_ID}:instruct": chat_template_kwargs: enable_thinking: false temperature: 1.0 top_p: 0.95 top_k: 64 min_p: 0.0 presence_penalty: 0.0 repeat_penalty: 1.0"