llama.cpp

9c35706b - CUDA: fix MMQ nwarps for AMD with warp_size==32 (#15014)