llama.cpp

gemma : more consistent attention scaling for v2 and v3

#13951

Merged