llama.cpp

Slight quantization improvement for Q4_K and Q5_K

#5361

Merged