Pytorch quite nonsensically supports rshift/lshift on non-integral inputs and shifts:
In [13]: a=torch.randn(4)
In [14]: a.bitwise_right_shift(2) #there's nothing bitwise about this shift
Out[14]: tensor([-0.0470, -0.0847, 0.0964, 0.0141])
In [15]: a.bitwise_right_shift(1.5) #what does this even mean?
Out[15]: tensor([-0.0664, -0.1198, 0.1363, 0.0200])
Pytorch quite nonsensically supports rshift/lshift on non-integral inputs and shifts:
numpy sensibly errors out in these cases
cc @ezyang
cc @mruberry @rgommers @nairbv