llama.cpp

Support requantizing models instead of only allowing quantization from 16/32bit

#1691

Merged