Likely bug in `tfp.math.fill_triangular_inverse`

I'm on tfp 0.16.0, python 3.8, windows10

Trying to extract a triangluar part of a matrix into a vector, I came across:

print(
    tfp.math.fill_triangular_inverse([
        [4, 0, 0],
        [6, 5, 0],
        [3, 2, 1]
    ])
)

print(
    tfp.math.fill_triangular_inverse([
        [4, 0, 0],
        [6, 5, 0],
        [3, 2, 1]
    ], upper=True)
)

which produces

tf.Tensor([1 2 3 4 5 6], shape=(6,), dtype=int32)
tf.Tensor([4 0 0 7 7 3], shape=(6,), dtype=int32)

and this makes absolutely no sense to me - where does the 7 come from and the 3 is clearly not part of the upper triangle....

Start by reproducing both tfp.math.fill_triangular_inverse calls from the issue with the shown matrix and upper=True. Check the documented ordering and expected upper-triangle contents, then trace the implementation responsible for the unexpected 7 values. Done means the upper-triangle result contains only the intended matrix elements and a regression test covers this example.