llama.cpp

bb07e28f - add --tensor-read-lazy