[sailingparrot]

Just for training and processing the existing context (pre fill phase). But when doing inference a token t has to be sampled before t+1 can so it’s still sequential