nit: self.reshape(-1) avoids some of the computation that goes on inside flatten.

Calling native::flatten instead of at::flatten inside of here means that we won't actually get autograd support. Using at::flatten (or reshape as pointed out above) fixes that problem. You can test this locally with something like:

x = torch.randn(3, 3, requires_grad=True)
out = x.ravel().sum()
out.backward()  # I think this throws an error with the current implementation

You are right. It doesn't really need to go into flatten.
But, t's interesting that it doesn't throw any error and gradient for the tensor is correct when I used native::flatten.

Oh, I see what happened now. native::flatten calls at::reshape, and at::reshape is supported by the autograd.

It's important to note here that "A copy is made only if needed."; that is part of the numpy semantics for ravel (https://numpy.org/doc/stable/reference/generated/numpy.ravel.html) and is also what the current implementation (via input.flatten()) does

nit: We should probably test more shapes. The implementation code for ravel is simple enough that I can believe it works for everything, but there's no harm in testing more

One edge case we should test is a tensor with a zero sized dimension. For example, the following should be true:

x = torch.randn([0, 2, 3]) # tensor of size [0, 2, 3]
y = x.ravel()
assert y.shape = ()