Y
Hacker News
new
|
ask
|
show
|
jobs
by
sailingparrot
172 days ago
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