Reproduce Gemma4 think loop with llama.cpp - View it on GitHub