Tests failing due to upstream OpenMDAO change related to check_partials error computation

[kejacobson]

An upstream change to OpenMDAO changes the way that errors are computed, and changes a name in error totals from forward_reverse to fwd_rev.

If we wish to maintain backwards compatibility of our tests with OpenMDAO, we'll need to check the version number of OpenMDAO in order to change the forward/reverse variable name as well as the tolerances we use for the finite differences since the tolerances used for previous versions of OpenMDAO now fail.

If we don't wish to maintain backwards compatibility of the tests, then we can change forward_reverse to fwd_rev and update the tolerances.

[kejacobson]

Rather than maintain two sets of tolerances, we're going to loosen the tolerance values to make both versions of error computation pass.

Direct usage of forward_reverse or fwd_rev will be refactored into a check_partials_output(error_data, tol) function that handles the version checking of OpenMDAO:

from distutils.version import LooseVersion
import openmdao

def check_partials_output(error_data, tol):
    assert_near_equal(error_data.reverse, 0.0, tolerance=tol)
    assert_near_equal(error_data.forward, 0.0, tolerance=tol)
    if LooseVersion(openmdao.__version__) <= LooseVersion("3.36.0"):
        assert_near_equal(error_data.forward_reverse, 0.0, tolerance=tol)
    else:
        assert_near_equal(error_data.fwd_rev, 0.0, tolerance=tol)

[Asthelen]

The forward_reverse and tolerance issues should be fixed in my network PR (#173).

The test_tacs_partials.py integration test now fails with OpenMDAO 3.36.1-dev but not 3.35.1-dev... it seems to be the forward partials of mass wrt x_struct0:

======================================================================
ERROR: test_derivatives (__main__.TestTACS) (partial_pair=('mass', 'x_struct0'))
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/lustre3/hpnobackup2/asthelen/Codes/Repositories/OM_test/mphys_network_2025jan24/tests/integration_tests/tmp/test_tacs_partials.py", line 111, in test_derivatives
    assert_near_equal(check_error.reverse, 0.0, 1e-7)
  File "/hpnobackup2/asthelen/Codes/Repositories/OM_test/OpenMDAO_2025jan24/openmdao/utils/assert_utils.py", line 724, in assert_near_equal
    raise ValueError('actual %s, desired %s, %s error %s, tolerance %s'
ValueError: actual [1.93536105], desired [0.], rel error 1.9353610478917778, tolerance 1e-07

I haven't checked the other two failing integration tests (with tacs/oas) but maybe the cause is the same?

[timryanb]

@Asthelen, I have no idea what's happening here. I can take a look

[timryanb]

@Asthelen @kejacobson It looks like OpenMDAO changed the relative derivative check to normalize on a element-by-element basis, rather than using the norm of the entire jacobian to normalize all errors. This leads to issues when we're checking against small near-zero terms in the Jacobian. Not sure how to get around this at the moment.

[timryanb]

Actually, this PR may help: OpenMDAO/OpenMDAO#3166

[kejacobson]

Closed in #204