llama.cpp

Use n_embd_gqa instead of n_embd to handle llama-2 70B

#2433

Merged