Most people are not even referring to CUDA batching nuances.

They think that sampling is an inherent part of Transformers.

Even on this site, it is regurgitated with confidence.