llama.cpp

8e6a9d2d - CUDA: more warps for mmvq on NVIDIA (#5394)