As reported on slack:

Can somebody explain the behavior of flip to me? I would expect that t.flip([a, b]) is a shorthand for t.flip(a).flip(b), but this is not the case (see below). Does the tuple variant flip along a diagonal or what?

In [12]: t = torch.arange(3*4*5).reshape(1, 3, 4, 5)

In [13]: t.flip([1, 3]).shape
Out[13]: torch.Size([5, 4, 3, 1])

In [14]: t.flip(1).flip(3).shape
Out[14]: torch.Size([1, 3, 4, 5])

ngimel says:

Looks like a bug. FWIW, flip on cuda tensor is not changing shape. There's some logic about permuting in the flip on cpu, that I don't quite get, may be it is to blame.
https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/TensorTransformations.cpp#L33-L54