llama.cpp

fbba272c - bench: add --tensor-read-lazy