llama.cpp

d052a0ed - Faster than CPU without 80% runtime memcpy