llama.cpp

8ace03ad - convert.py : better always have n_head_kv and default it to n_head