Failures from MINLPTests

[odow]

I'm running Uno over our standard test-set: jump-dev/AmplNLWriter.jl#190

I'll extract some reproducible examples of failures for you to investigate.

Actually, I think all these failures stem from the same problem. The first comment has the simplest reproducible example.

The error appears to be The inertia correction got unstable (delta_w > threshold)

[cvanaret]

A quick comment on the first problem: I can solve the generated nl file on my machine, so I cannot explain the error yet.
That said, the nl file is incorrectly generated: it contains two variables and one constraint, instead of one variable (and one bound constraint if you interpret it correctly). Nonetheless, Uno can solve the original 1D problem as well.

[odow]

JuMP interprets @constraint(model, x >= 2) as a linear constraint. Bounds are given as @variable(model, x >= 2).

I see only one variable in the NL file? Which part is wrong? I get so confused with nl files

[odow]

I can solve the generated nl file on my machine

Is this some MUMPS version issue then? What version of MUMPS are you compiling with?

[odow]

How do I turn on maximally verbose logs for you?

[cvanaret]

JuMP interprets @constraint(model, x >= 2) as a linear constraint. Bounds are given as @variable(model, x >= 2).
I see only one variable in the NL file? Which part is wrong? I get so confused with nl files

Ah! In this case, I see the dimensions of the problem reformulated by Uno and it had to introduce a slack variable. This is not great but at least expected. All good!

Is this some MUMPS version issue then? What version of MUMPS are you compiling with?

I'm using MUMPS 5.7.2.

How do I turn on maximally verbose logs for you?

That's a good idea. Pass the value DEBUG3 to the option logger. Thanks!

[odow]

I'm not sure what changed. I'll re-run the MINLPTests.

[cvanaret]

Wonderful :)

[odow]

Okay, I think I've figured this out: jump-dev/AmplNLWriter.jl#190 (comment)

The tests were using a slightly different version of MUMPS to the one we compiled with. We can fix this in the next build by adding strict compat bounds.

This is one issue with our approach of using shared libraries... Even though MUMPS v5.6 should be compatible with v5.7, the evidence clearly shows that it isn't...

[odow]

Closing because this is fixed and was a mistake on my part. The issue was the difference in MUMPS version.

I've opened #38 and #39 with some of the remaining issues.