llama.cpp

3f620574 - llama: properly handle KV on training