llama.cpp

Allow quantizing k-quants to fall back when tensor size incompatible

#3747

Merged