llama.cpp

a710d58d - Try fix quantized k-cache on ROCm