Reproducing code example:

svd with hermitian=True (introduced by @eric-wieser in #12693) has a bug with the order of singular values:

import numpy as np

a = np.array([[0, 1, 0], [1, 0, 1], [0, 1, 0]])
u, s, v = np.linalg.svd(a, hermitian=True)
print(s)

Outputs: [1.41421356e+00 6.07153217e-18 1.41421356e+00]

Per the contract of svd, the singular values should be sorted in descending order, but they aren't here. That's because we simply reverse the order of eigenvalues than sorting, which only suffices if all eigenvalues happen to be positive.

Numpy/Python version information:

1.17.5 3.6.9 (default, Nov 7 2019, 10:44:02)
[GCC 8.3.0]