llama.cpp

Make IQ4_NL quantization be the same on CPU/CUDA/Metal when quantizing K-cache

#6196

Merged