llama.cpp

d29fb057 - metal : take into account the actual allocated memory of the tensor