Hacker News new | past | comments | ask | show | jobs | submit
The only scenario is if you have enough work to do batch inference. Using a tiny fraction of GPU capacity to decode a single request at a time just doesn't make sense, as you say.