llama.cpp

0bf56369 - convert : yield Gemma3N custom_map tensors directly (#19091)