Ang Wang

Unexpected performance of dropout

1

hi, recently I've been doing some torch xla performance analysis on gpu.  I came across with a problem that the default random algorithm with poor performance. I use bert base...

As described in https://github.com/pytorch/xla/issues/6589 and https://github.com/pytorch/xla/issues/6294, re-implementing the lowering of aten::sub and aten::rsub. @wonjoolee95 @JackCaoG