Autoregressive transformers take a while to generate text, since you need to run the whole model once for every word in the output.