llama.cpp

396856b4 - CUDA/HIP: add support for selectable warp size to mmv (#11519)