This workaround logic in llama.cpp is causing crashes for users with less system memory than VRAM.
llama