llama.cpp

0fc1e820 - CUDA: faster large batch FA without tensor cores (#7314)