Created by: chenwhql

PR types

Bug fixes

PR changes

Others

Describe

Reverts PaddlePaddle/Paddle#24895

• revert reason:

After cub was upgraded to version 1.9.8, a single training process will occupy all currently visible GPUs, which has two effects:

1. training start speed is significantly slower
2. make the experience of using multi-process multi-cards training worse

There may be bugs in this version of cub, so revert this commit