llama.cpp

d5c8fcfd - convert.py : 70b model working (change attn_q permute)