llama.cpp

a20f3c74 - CUDA: fix tensor core logic for Pascal and HIP (#4682)