Excluded tests

[odow]

It's be nice to know why these tests are failing:

cuOpt.jl/test/MOI_wrapper.jl

Line 53 in a67b6d1

exclude = ["test_constraint_ZeroOne_bounds_3", "test_solve_TerminationStatus_DUAL_INFEASIBLE", "test_unbounded_MAX_SENSE", "test_unbounded_MIN_SENSE"],

test_constraint_ZeroOne_bounds_3

This is a trivial test that I expect you to pass

https://github.com/jump-dev/MathOptInterface.jl/blob/703a15e3d5b4bb6b8040bfd07540e2fe4c4976f3/src/Test/test_constraint.jl#L617-L633

test_solve_TerminationStatus_DUAL_INFEASIBLE

This one is up for debate, but if you can't pass it, I'm surprised if you can detect any unboundedness?

https://github.com/jump-dev/MathOptInterface.jl/blob/703a15e3d5b4bb6b8040bfd07540e2fe4c4976f3/src/Test/test_solve.jl#L200-L219

test_unbounded_MAX_SENSE and test_unbounded_MIN_SENSE

I see that you claim the primal result is an infeasibility certificate:

cuOpt.jl/src/MOI_wrapper.jl

Line 461 in a67b6d1

return MOI.INFEASIBILITY_CERTIFICATE

It it actually a ray? If so, it must follow this convention: https://jump.dev/MathOptInterface.jl/stable/background/infeasibility_certificates/#Unbounded-problems

It it isn't a ray and is just the last feasible iterate, the PrimalStatus should be FEASIBLE_POINT.

[chris-maes]

@rg20 is out of office. So I don't have the full context here. I'll try to answer as best I can.

I'm not sure about test_constraint_ZeroOne_bounds_3. @rgsl888prabhu maybe you can enable it and post the logs and we can see why it is failing.

For test_solve_TerminationStatus_DualInfeasible, we can detect if a problem is unbounded with dual simplex. But we may struggle with PDLP. I'm guessing that is why the test is disabled.

For test_unbounded_MAX_SENSE and test_unbounded_MIN_SENSE, again we are able to detect if a problem is unbounded in dual simplex from the ratio test. It is possible to convert this to a Farkas ray. But we don't currently provide a Farkas ray or an infeasibility certificate at this time.

Do you have the ability to declare a problem unbounded without providing an unbounded ray? If so, this seems preferable than returning a 'FEASIBLE_POINT' status.

[odow]

test_solve_TerminationStatus_DualInfeasible

The problem is:

model = Model()
@variable(model, x[1:5])
@objective(model, Min, sum(x))
optimize!(model)

test_unbounded_MAX_SENSE

Do you have the ability to declare a problem unbounded without providing an unbounded ray?

Yes. There is MOI.TerimationStatus that will return MOI.DUAL_INFEASIBLE. There is a separate MOI.PrimalStatus that can return one of MOI.INFEASIBILITY_CERTIFICATE, MOI.FEASIBLE_POINT, or MOI.NO_SOLUTION, depending on how the solver finished.

For now I've just removed the incorrect MOI.INFEASIBILITY_CERTIFICATE: #12

If the solver always has a feasible point when it proves unboundedness, then we could return MOI.FEASIBLE_POINT instead.

If by "unbounded" you sometimes mean dual infeasible without checking primal feasibility and thus don't have a primal feasible point, we can stick with returning MOI.NO_SOLUTION.

[rgsl888prabhu]

And regarding test_constraint_ZeroOne_bounds_3 I have created an issue NVIDIA/cuopt#260

[chris-maes]

If by "unbounded" you sometimes mean dual infeasible without checking primal feasibility and thus don't have a primal feasible point, we can stick with returning MOI.NO_SOLUTION.

Yes, we don't currently check for a primal feasible point when returning this status. We can try to clean up our status codes a bit to distinguish between Dual Infeasible and Primal Unbounded.

[odow]

Yes, we don't currently check for a primal feasible point when returning this status

To clarify: you don't check for primal feasibility before returning CUOPT_TERIMINATION_STATUS_UNBOUNDED?

This is why JuMP does not have an explicit UNBOUNDED termination status. We have only DUAL_INFEASIBLE, and primal unboundedness is confirmed either with the presence of a primal FEASIBLE_POINT, or the user must resolve the model without an objective.

[odow]

Closing. These are tracked upstream:

cuOpt.jl/test/MOI_wrapper.jl

Lines 58 to 62 in 81b8fbf

	exclude = [
	# Upstream bug: https://github.com/NVIDIA/cuopt/issues/260
	"test_constraint_ZeroOne_bounds_3",
	# Upstream bug: https://github.com/NVIDIA/cuopt/issues/112
	"test_solve_TerminationStatus_DUAL_INFEASIBLE",