This includes the size of the Storage in the tensor, but the documentation for sizeof says:

Only the memory consumption directly attributed to the object is accounted for, not the memory consumption of objects it refers to.

That means that the recipe for computing memory referenced in the Python documentation will overstate the memory usage:

https://code.activestate.com/recipes/577504/

I don't think this will change the code snippet from the original PyTorch discussion either:
https://discuss.pytorch.org/t/data-type-when-change-tensor-array-to-numpy-array/9786

test0 = torch.rand(100, 100)
test1 = test0.numpy()
print(sys.getsizeof(test1))

test1 is a NumPy array. It's size is reported as 112 because it's a view and test1.base is intentionally not included in __sizeof__.

On the other hand, perhaps including the Storage size in the Tensor is useful enough for debugging that it's not worth trying to strictly adhere to Python documentation spec.

Thank you @colesbury. I'm good with sticking to the specification.
However storage is so seldom used explicitly in PyTorch, contrarily to Lua Torch, that it feels like the analogous of the data buffer in NumPy.

Done