DeepSpeed
69a45988 - Stage_1_and_2.py: fix assert for reduce_scatter configurations combinations (#4964)

Commit
1 year ago
Stage_1_and_2.py: fix assert for reduce_scatter configurations combinations (#4964) today it does not allow to set reduce_scatter with: - invalid comm data type - gradient_predivide_factor != 1.0 - postscale_gradients is False but reduce scatter is relevant only for Zero2, also reduce_scatter is set to True by default. So setting prescale_gradients=True in json will end up with assert on Zero1. Co-authored-by: Olatunji Ruwase <olruwase@microsoft.com>
Author
Parents
Loading