llama.cpp

c2beeb8e - only allocate as much memory as is required in each backend for the model