llama.cpp

1.5 bit: we can do even better

#5999

Merged