llama.cpp

b64d7cc2 - cuda: unary ops as float + de-duplicate (ggml/1130)