Support QP interface via MOI by mtanneau · Pull Request #30 · jump-dev/cuOpt.jl

mtanneau · 2026-01-10T20:21:02Z

This PR introduces support for solving Quadratic Programs with cuOpt.

Notes:

cuOpt only supports QPs with convex objectives. The MOI wrapper does not perform any convexity checks, it passes the matrix as-is to the solver
while building the wrapper, I encountered an upstream bug (see [BUG] cuOpt crashes when solving QP with no linear constraints NVIDIA/cuopt#759) that led me to disable two MOI tests
I have tested this locally on a DGX Spark machine as follows:
- OS: Ubuntu (Linux aarch64)
- GPU: GB10
- cuOpt: v25.12.0

Please see extra comments in the diff

mtanneau · 2026-01-10T20:24:22Z

+    # Is this a QP or an LP?
+    has_quadratic_objective = length(qobj_matrix_values) > 0
+    if has_quadratic_objective && has_integrality
+        error("cuOpt does not support models with quadratic objectives _and_ integer variables")


Not sure what the best error / error message is here, happy to modify this as requested.

Note that cuOpt does not support MIQPs, only MILP or continuous QP

This error is okay for now.

As other options: you could remove the error and make it so that TerminationStatus was MOI.INVALID_MODEL. Or you could throw(MOI.SetAttributeNotAllowed(obj_attr, "cuOpt does not support ...")) but it's a coin toss whether the objective or integer constraint is the problem. And MOI doesn't have a bridge to fix this so the error is going to propagate to the user anyway.

mlubin · 2026-01-11T01:45:37Z

From the README:

Note: This version of cuOpt.jl supports the Nvidia cuOpt 25.08, 25.10, and 25.12 releases.

Which version of cuOpt introduced this QP API?

mtanneau · 2026-01-11T03:46:39Z

Which version of cuOpt introduced this QP API?

It was added to the C API in cuOpt v25.12 (See release notes)

mlubin · 2026-01-11T17:23:33Z

@rgsl888prabhu should weigh in on if we want to bump the minimum supported version of cuOpt to 25.12 or add appropriate error messages when users try to solve QPs with an older version. Bumping the required version is ok with me since it makes maintenance of the wrapper easier.

blegat · 2026-01-12T07:40:19Z

I think DCO is failing because the commit made via the github UI were not signed off

mtanneau · 2026-01-12T14:54:45Z

I think DCO is failing because the commit made via the github UI were not signed off

Thank you, fixed and rebased

rg20

Thank you for implementing the QP interface for cuOpt!

I think the cuOpt version support might need changes in cuOpt.jl file. The QP support only works from 25.12.

rg20 · 2026-01-12T13:58:18Z

+        v = qterm.coefficient
+        if i == j
+            # Adjust diagonal coefficients to match cuOpt convention
+            v /= 2


Not sure if this is accurate.

cuOpt expects users to provide the true objective function.

For example: if you are minimizing x1^2 + x2^2, the matrix should be [1 0; 0 1], cuOpt internally minimizes for (1/2) [x1 x2] [2 0; 0 2] [x1; x2] in this case.

cuOpt expects users to provide the true objective function.

https://www.youtube.com/watch?v=M31xoZGyj9w&t=2698s

What happens to the off-diagonal terms?

cuOpt is not necessarily expecting the matrix to be symmetric.

If the objective is xT Q x + cT x, we internally symmetrize and solve for (1/2) xT (Q + QT) x + cT x

So if the objective is x1^2 + x2^2 + 2 x1 x2,

Q can be: [1 2; 0 1], or [1 0; 2 1] or [1 1; 1; 1]

(including the links for completeness)

cuOpt docs: https://docs.nvidia.com/cuopt/user-guide/latest/lp-qp-features.html#quadratic-programming

MOI docs: https://jump.dev/MathOptInterface.jl/stable/reference/standard_form/#MathOptInterface.ScalarQuadraticFunction

I believe the /2 factor for diagonal terms is required for correctness, but agreed it's not super clear why at first.

The current MOI wrapper uses a JuMP-level data structure to store the optimization model as the user builds it. That internal representation abides by MOI conventions.

when optimize! gets called, the problem data is flushed from the JuMP-level data store into a cuOpt-level data structure --> this is why this PR mostly modifies the copy_to function.

Hence, when we access the quadratic objective in copy_to, we get as input an MOI.ScalarQuadraticFunction, which abides by the MOI convention described in the docs above.

Here is a small example to illustrate this

using JuMP model = Model() @variable(model, x) @variable(model, y) @objective(model, Min, x*x + 2*x*y + 3*y*y) objective_function(model) # x² + 2 x*y + 3 y² # Now, access the MOI representation F = MOI.get(model, MOI.ObjectiveFunctionType()) f = MOI.get(model, MOI.ObjectiveFunction{F}()) f.quadratic terms

the last line outputs

3-element Vector{MathOptInterface.ScalarQuadraticTerm{Float64}}: MathOptInterface.ScalarQuadraticTerm{Float64}(2.0, MOI.VariableIndex(1), MOI.VariableIndex(1)) MathOptInterface.ScalarQuadraticTerm{Float64}(2.0, MOI.VariableIndex(1), MOI.VariableIndex(2)) MathOptInterface.ScalarQuadraticTerm{Float64}(6.0, MOI.VariableIndex(2), MOI.VariableIndex(2))

--> you can see that the diagonal coefficients were multiplied by 2. This was done by MOI in accordance with its convention.

@rg20 I can add the following:

add link to the MOI docs on quadratic function in that comment, to provide additional context

add a couple of unit tests to validate the cuOpt solution and objective value when solving QP from MOI

My conclusion with this all is that the only valid way is to test, test, test, test, and test. There are too many subtleties to try and logically reason about the transformations, and every time I do, I end up making a mistake.

There are tests in MOI for various cases, but these are precisely the ones that you're skipping because of the upstream bug... 😢 ("test_objective_qp_ObjectiveFunction_zero_ofdiag" and "test_objective_qp_ObjectiveFunction_edge_cases").

@mtanneau thanks for explaining the subtleties in the MOI wrapper.

Added more unit tests that specifically trigger the QP objective.
Notes:

the upstream issue regarding QP with no constraints might get fixed in cuopt v26.02 (tracking Use augmented system when there are no constraints NVIDIA/cuopt#765), which would allow to run more MOI-level tests.

if the team is OK with adding JuMP as a test dependency, I'm happy to add similar tests where the QP is built from JuMP, not from MOI.

JuMP shouldn't be a test dependency; MOI-style tests should be sufficient

rgsl888prabhu · 2026-01-12T20:29:05Z

@rgsl888prabhu should weigh in on if we want to bump the minimum supported version of cuOpt to 25.12 or add appropriate error messages when users try to solve QPs with an older version. Bumping the required version is ok with me since it makes maintenance of the wrapper easier.

@mlubin Yes, lets bump it to 25.12.

rgsl888prabhu · 2026-03-03T03:54:04Z

Shall we bump this 26.02 and merge this PR ?

mtanneau · 2026-03-03T03:59:10Z

Bumping to 26.02 might need to update the C API bindings. Should be straightforward, happy to give it a try tomorrow (ie Tuesday for me).

…

On Mon, Mar 2, 2026 at 22:54 Ramakrishnap ***@***.***> wrote: *rgsl888prabhu* left a comment (jump-dev/cuOpt.jl#30) <#30 (comment)> Shall we bump this 26.02 and merge this PR ? — Reply to this email directly, view it on GitHub <#30 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACJGBQM33H3VIPNRHKVX5QD4OZJPFAVCNFSM6AAAAACRJKLK2CVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTSOBYGQ4DAOJQGQ> . You are receiving this because you were mentioned.Message ID: ***@***.***>

mtanneau · 2026-03-03T22:00:05Z

Re-v26.02: I opened a separate issue (#32) and will open a separate PR for the C API and existing MOI interface.

I'll keep all the QP API wrapping in MOI in this PR, so will come back here once #32 is addressed.

rgsl888prabhu · 2026-03-04T02:35:45Z

@odow @rg20 shall we merge this ?

odow · 2026-03-04T02:50:15Z

No opinion. I assume it is passing tests?

rgsl888prabhu · 2026-03-04T03:02:26Z

@mtanneau can you please confirm all tests are passing ?

mtanneau · 2026-03-04T15:57:29Z

I just re-ran tests locally on this branch, building with cuOpt v26.02. Putting aside the same errors that I encountered in #33 (i.e., upstream issue in cuOpt where dual infeasibility is detected at presolve), tests are passing.

I would rather merge #33 first, so that the cuOpt compat requirements matches the C API bindings.
This branch sets cuopt compat to v26.02, but uses C API bindings that were generated from v25.XX, which feels off (despite working seemingly well).

mlubin · 2026-04-16T19:38:32Z

@mtanneau please feel free to rebase this and we'll get it reviewed and landed quickly.

Signed-off-by: mtanneau <mathieu.tanneau@gmail.com>

mtanneau · 2026-04-19T21:31:55Z

rebased and updated!
Tests are passing locally with cuOpt v26.04

mlubin · 2026-04-20T00:22:03Z

+        0.25;
+        atol = 1e-4,
+        rtol = 1e-4,
+    )


Feels like a good idea to test the dual solutions as well. Note you might run into NVIDIA/cuopt#1119.

Why aren't the MOI tests hitting this?

I don't know if NVIDIA/cuopt#1119 is a bug in cuopt or just in the cvxpy wrapper. If it's the latter then the MOI tests won't hit it.

Currently the MOI interface doesn't appear to support querying MOI.ConstraintDual attribute, so I would suggest holding off on that and adding support in a separate PR

odow · 2026-04-20T00:27:19Z

+        MOI.ScalarAffineFunction(
+            [MOI.ScalarAffineTerm(1.0, x), MOI.ScalarAffineTerm(1.0, y)],
+            0.0,
+        ),


You can use operator overloading to make the functions: 1.0 * x + 1.0 * y. It makes things a lot clearer.

odow · 2026-04-20T00:27:50Z

+        ],
+        MOI.ScalarAffineTerm{Float64}[],
+        0.0,
+    )


Ditto here: 0.5 * x * x + 0.5 * y * y. Then there's no confusion with the 1/2

odow · 2026-04-20T00:29:12Z

+        ],
+        [MOI.ScalarAffineTerm(-4.0, x), MOI.ScalarAffineTerm(+4.0, y)],
+        4.0,
+    )


Strongly prefer to create new test functions rather than having long complicated tests.

You can test:

a diagonal QP objective

an off-diagonal QP objective

changing from QP to LP

etc

Do you even need these tests? They should be covered by one the ones in MOI

Good point, I wrote those tests to

circumvent some of the excluded MOI tests and/or upstream issues,

validate the MOI vs cuOpt convention regarding QP objective (that infamous 1/2 coefficient)

I removed them as they are indeed covered by MOI tests.

Signed-off-by: mtanneau <mathieu.tanneau@gmail.com>

Signed-off-by: Miles Lubin <mlubin@nvidia.com>

mtanneau commented Jan 10, 2026

View reviewed changes

Comment thread Project.toml

mtanneau commented Jan 10, 2026

View reviewed changes

Comment thread src/MOI_wrapper.jl

mtanneau commented Jan 10, 2026

View reviewed changes

Comment thread src/MOI_wrapper.jl Outdated

mtanneau commented Jan 10, 2026

View reviewed changes

mlubin reviewed Jan 10, 2026

View reviewed changes

Comment thread src/MOI_wrapper.jl Outdated

mtanneau force-pushed the qp-interface-moi branch from 64e7474 to 3a13f8c Compare January 10, 2026 22:41

mlubin reviewed Jan 11, 2026

View reviewed changes

Comment thread src/MOI_wrapper.jl Outdated

Comment thread src/MOI_wrapper.jl Outdated

Comment thread src/MOI_wrapper.jl Outdated

odow reviewed Jan 11, 2026

View reviewed changes

Comment thread src/MOI_wrapper.jl Outdated

mtanneau force-pushed the qp-interface-moi branch from c1313e4 to 067d973 Compare January 12, 2026 14:52

rg20 reviewed Jan 12, 2026

View reviewed changes

mtanneau added 6 commits April 18, 2026 22:46

Support QP interface via MOI

b9baa2d

Signed-off-by: mtanneau <mathieu.tanneau@gmail.com>

Move objective data extraction to separate method

efec7d5

Signed-off-by: mtanneau <mathieu.tanneau@gmail.com>

Format

9e9cc63

Signed-off-by: mtanneau <mathieu.tanneau@gmail.com>

Fix typos

5141b73

Signed-off-by: mtanneau <mathieu.tanneau@gmail.com>

In-place offset + remove comments

41822fe

Signed-off-by: mtanneau <mathieu.tanneau@gmail.com>

Move import

9a66093

Signed-off-by: mtanneau <mathieu.tanneau@gmail.com>

mtanneau added 3 commits April 18, 2026 22:46

Explain MOI vs cuOpt convention re:quadratic functions

ef21583

Signed-off-by: mtanneau <mathieu.tanneau@gmail.com>

Add QP unit tests

1dfe699

Signed-off-by: mtanneau <mathieu.tanneau@gmail.com>

Fix typos + format

335ce40

Signed-off-by: mtanneau <mathieu.tanneau@gmail.com>

mtanneau force-pushed the qp-interface-moi branch from cba594b to 335ce40 Compare April 19, 2026 21:20

mlubin reviewed Apr 20, 2026

View reviewed changes

odow reviewed Apr 20, 2026

View reviewed changes

Remove QP tests coverd by MOI

be427f7

Signed-off-by: mtanneau <mathieu.tanneau@gmail.com>

mlubin merged commit 706afa4 into jump-dev:main Apr 21, 2026
3 checks passed

mlubin referenced this pull request Apr 21, 2026

Update Project.toml for 0.2.1 release (#36)

93dc73c

Signed-off-by: Miles Lubin <mlubin@nvidia.com>

mtanneau mentioned this pull request Apr 22, 2026

Add support for MOI.ConstraintDual #37

Open

Conversation

mtanneau commented Jan 10, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

odow Jan 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

mlubin commented Jan 11, 2026

Uh oh!

mtanneau commented Jan 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mlubin commented Jan 11, 2026

Uh oh!

Uh oh!

blegat commented Jan 12, 2026

Uh oh!

mtanneau commented Jan 12, 2026

Uh oh!

rg20 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mtanneau Jan 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

rg20 Jan 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

rgsl888prabhu commented Jan 12, 2026

Uh oh!

rgsl888prabhu commented Mar 3, 2026

Uh oh!

mtanneau commented Mar 3, 2026 via email

Uh oh!

mtanneau commented Mar 3, 2026

Uh oh!

rgsl888prabhu commented Mar 4, 2026

Uh oh!

odow commented Mar 4, 2026

Uh oh!

rgsl888prabhu commented Mar 4, 2026

Uh oh!

mtanneau commented Mar 4, 2026

Uh oh!

mlubin commented Apr 16, 2026

Uh oh!

mtanneau commented Apr 19, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

odow Jan 12, 2026 •

edited

Loading

mtanneau commented Jan 11, 2026 •

edited

Loading

mtanneau Jan 12, 2026 •

edited

Loading

rg20 Jan 13, 2026 •

edited

Loading

mtanneau Apr 21, 2026 •

edited

Loading