llama.cpp

0f2e42ca - CUDA: only allocate FA tmp buffer if needed (#18564)