llama.cpp

b9e92c43 - CUDA: dequant quantized kv on-the-fly for vec kernels