llama.cpp

0aac18be - use memcpy for small copies across host-visible memory