Test test_gmres_on_larger_random_problem fails without seeds

[crawlingcub]

Hi,

The test test_gmres_on_larger_random_problem in tensornetwork/backends/numpy/numpy_backend_test.py link always fails (500 fails out of 500 runs) for [float16] parameter.

The interesting part here is that the err and max(rtol, atol) are always exactly similar without seeds (as shown below)! Do you think this test needs a higher tolerance than what is being used when testing for float16? Or is there a deeper bug?

Err message:

>     assert err < max(rtol, atol)
E     assert 11.87 < 9.3994140625
E      +  where 9.3994140625 = max(9.3994140625, 0.09765625)