Description
PyTorch before 3.7.0 has a bernoulli_p decompose function in decompositions.py even though it lacks full consistency with the eager CPU implementation, negatively affecting nn.Dropout1d, nn.Dropout2d, and nn.Dropout3d for fallback_random=True.
References
github.com/pytorch/pytorch/issues/142853
github.com/pytorch/pytorch/pull/143460
gist.github.com/shaoyuyoung/e636f2e7a306105b7e96809e2b85c28a
github.com/pytorch/pytorch/compare/v2.6.0...v2.7.0
gist.github.com/shaoyuyoung/4bcefba4004f8271e64b5185c95a248a