llama.cpp

llama: properly handle KV on training

#28520

Open