llama.cpp

11f0af55 - CUDA: faster tile FA, add oob checks, more HSs (#16492)