llama.cpp

63b64a50 - metal : dequant kv cache only for large batches (#27438)