zero divide when using grad rho

[DLWoodruff]

Ryan is getting zero divide in this code:

            rhos = np.fromiter(
                (
                    abs(costs[s][ndn, i]/loc_denom[s][ndn, i])
                    for i, v in enumerate(node.nonant_vardata_list)
                ),
                dtype="d",
                count=nlens[ndn],
            )

File "/storage/project/r-dmolzahn6-0/rpiansky3/venvs/ph_batts_py310/src/mpi-sppy/mpisppy/extensions/grad_rho.py", line 211, in
abs(costs[s][ndn, i]/loc_denom[s][ndn, i])

Unless someone has a better idea, I am inclined to do an out=np.xxxx with a where=loc_denom[s][ndn, i] != 0

for xxxx, I am inclined to use np.ones(1), but I can see how one might think a large value makes more sense.

[bknueven]

Do you know if this was run with the "dependent" or "independent" denominator?

I am guessing "independent", in which case maybe we add the same logic to the independent denominator as we have in the dependent denominator:

mpi-sppy/mpisppy/extensions/grad_rho.py

Lines 62 to 67 in 459cafe

	denom_max = max(scen_dep_denom.values())
	for ndn_i, v in s._mpisppy_data.nonant_indices.items():
	if (scen_dep_denom[ndn_i]) <= self.denom_bound * v._value:
	# denom_max is over all nonants, so if not >0 we will stop
	scen_dep_denom[ndn_i] = max(denom_max, self.denom_bound * v._value)