hard20 pts · first answer countstransformerspositional-encodinginference

Left padding with absolute positions

You left-pad a batch for a model that uses position ids, but you compute position ids as 0, 1, 2, ... from the start of each padded row. What problem does this cause?

A.Real tokens get shifted positions that depend on how much padding the row hasB.The padding tokens are generated as part of the answer textC.The attention mask stops working for all rows in the batchD.The model's vocabulary is reduced by the number of pad tokens

Discussion 0

Join free to post in the discussion.

Left padding with absolute positions

Discussion 0

Related questions