Store best dual objective, align root LP bound with relative_mip_gap by anandhkb · Pull Request #957 · NVIDIA/cuopt

anandhkb · 2026-03-14T17:04:04Z

Callback: store only best (tightest) dual objective; it can worsen during iterations (min: keep largest, max: keep smallest)
set_final_solution: when returning with -inf, push root_lp_current_lower_bound_ to stats so relative_mip_gap uses the best/tightest bound
user_bound_callback: only update stats when new bound is better, avoid overwriting e.g. set_simplex_solution with a worse root LP value
Log line change: 'Root LP dual objective (last)' -> '(best)'

- Callback: store only best (tightest) dual objective; it can worsen during iterations (min: keep largest, max: keep smallest) - set_final_solution: when returning with -inf, push root_lp_current_lower_bound_ to stats so relative_mip_gap uses the best bound - user_bound_callback: only update stats when new bound is better, avoid overwriting e.g. set_simplex_solution with worse root LP value - Log: 'Root LP dual objective (last)' -> '(best)'

copy-pr-bot · 2026-03-14T17:04:08Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

coderabbitai · 2026-03-14T17:08:30Z

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: d5b3d0be-ba67-4e88-8506-bdd5e69f20b3

📥 Commits

Reviewing files that changed from the base of the PR and between d531ad1 and 6d1919e.

📒 Files selected for processing (2)

cpp/src/branch_and_bound/branch_and_bound.cpp
cpp/src/mip_heuristics/solver.cu

📝 Walkthrough

Walkthrough

The changes improve bound tracking and reporting in optimization solvers by introducing objective-sense-aware bound comparisons. The dual objective callback now only updates bounds when they represent improvements relative to the optimization direction, and messaging clarifies that reported bounds are the best observed. Early termination scenarios properly propagate updated bounds to user callbacks.

Changes

Cohort / File(s)	Summary
Bound tracking refinement `cpp/src/branch_and_bound/branch_and_bound.cpp`	Root LP dual objective callback now stores bounds only when they represent the best (tightest) bound accounting for objective sense (maximization vs. minimization). Log messaging updated to reflect "best" rather than "last" bound. Early-exit scenarios with -inf lower bound now propagate finite root bounds to user-bound callbacks.
Solution-bound callback improvement `cpp/src/mip_heuristics/solver.cu`	Simple solution-bound update replaced with objective-sense-aware logic. Introduces is_minimization check and computes bound improvements based on current bound finiteness and optimization direction, updating stored bound only when the new bound is superior.

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

🚥 Pre-merge checks | ✅ 2 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately summarizes the main changes: storing the best dual objective and aligning the root LP bound with relative_mip_gap calculation.
Description check	✅ Passed	The description directly relates to the changeset, detailing the callback behavior changes, set_final_solution modifications, user_bound_callback updates, and logging changes reflected in the code modifications.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create PR with unit tests
Post copyable unit tests in a comment

📝 Coding Plan

Generate coding plan for human review comments

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

Tip

You can get early access to new features in CodeRabbit.

Enable the early_access setting to enable early access features such as new models, tools, and more.

anandhkb · 2026-03-14T18:44:43Z

/ok to test 6d1919e

chris-maes · 2026-03-16T16:41:20Z

cpp/src/branch_and_bound/branch_and_bound.cpp

  lp_settings.concurrent_halt                 = get_root_concurrent_halt();
  lp_settings.dual_simplex_objective_callback = [this](f_t user_obj) {
-    root_lp_current_lower_bound_.store(user_obj);
+    // Store only the best (tightest) dual objective; it can worsen during iterations


The dual objective should always increase over the iterations.

Ok, thanks, I refer to Burcin's testcase:

cuopt_cli --log-file Optimality_MIP_1.log --solution-file Optimality_MIP_1.sln --time-limit 28500 Optimality_MIP_1.mps

Issue is: The dual obj value is getting worse rather than monotonically improving. He says - "If you have taken the last dual obj for your gap calculation, the resulting value is not correct because a better dual obj value was found in an earlier iteration". Here is the log Burcin referred me to:

236000 -7.8615507602829341e+04 33078 2.15752325e+16 0.00e+00 873.90 237000 -7.8599148858700210e+04 32968 2.33309309e+16 0.00e+00 878.39 238000 -7.8628596397306974e+04 31778 4.13809980e+15 0.00e+00 882.75 239000 -7.8625353538739626e+04 31793 4.09842601e+15 0.00e+00 886.47 240000 -3.4223435428449867e+05 34910 4.33595251e+22 0.00e+00 894.38. ### this is where dual obj value gets worse 241000 -4.9084521357962309e+05 33468 3.08252689e+20 0.00e+00 899.81. ### and worse... 242000 -5.2980592629211734e+05 33705 7.28784267e+19 0.00e+00 904.61 243000 -5.3195161855784897e+05 32822 7.03858147e+18 2.00e-07 908.23 244000 -5.3188925240266300e+05 32597 2.46222227e+19 0.00e+00 912.64 ... 279000 -5.7260862863620499e+05 24915 1.06134033e+18 0.00e+00 1234.99 280000 -5.7271001072192355e+05 24276 7.90538350e+17 0.00e+00 1241.53 281000 -5.7235100939733768e+05 24464 1.18694208e+18 0.00e+00 1257.79 New solution from primal heuristics. Objective -5.993702e+04. Gap 854.9%. Time 1258.26. ### a heuristic solution found but gap should be 31.18%, in relation to line 239000 282000 -5.7285144600908319e+05 20031 1.07701601e+13 1.00e-07 1276.96 283000 -5.7284553597028111e+05 23587 4.28140763e+14 0.00e+00 1287.60

That looks like an issue in either numerics or perturbation in dual simplex. We should fix the issue inside of dual simplex, rather than adding code to hide the issue outside of dual simplex.

Please let me know how to reproduce the above log.

chris-maes

Is this PR necessary?

anandhkb · 2026-03-16T20:26:14Z

@bbozkaya and @chris-maes to discuss further and identify IF there is an actual need for this PR. Thanks @chris-maes and @bbozkaya !

anandhkb · 2026-03-16T21:07:00Z

Attaching the cuopt log file (partial) for the problem on which @bbozkaya saw an anomaly. I ran from a build based off main branch today.
cuopt_log.txt

chris-maes · 2026-03-20T15:27:53Z

I will close this for now. I think this will require a larger effort to address.

anandhkb added this to the 26.04 milestone Mar 14, 2026

anandhkb requested a review from a team as a code owner March 14, 2026 17:04

anandhkb added non-breaking Introduces a non-breaking change improvement Improves an existing functionality labels Mar 14, 2026

anandhkb requested review from akifcorduk and chris-maes March 14, 2026 17:04

chris-maes reviewed Mar 16, 2026

View reviewed changes

chris-maes requested changes Mar 16, 2026

View reviewed changes

rgsl888prabhu changed the base branch from main to release/26.04 March 19, 2026 16:29

chris-maes closed this Mar 20, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Store best dual objective, align root LP bound with relative_mip_gap#957

Store best dual objective, align root LP bound with relative_mip_gap#957
anandhkb wants to merge 1 commit intoNVIDIA:release/26.04from
anandhkb:followUpToLogPR

anandhkb commented Mar 14, 2026

Uh oh!

copy-pr-bot bot commented Mar 14, 2026

Uh oh!

coderabbitai bot commented Mar 14, 2026

Walkthrough

Changes

Estimated code review effort

❌ Failed checks (1 warning)

Uh oh!

anandhkb commented Mar 14, 2026

Uh oh!

chris-maes Mar 16, 2026

Uh oh!

anandhkb Mar 16, 2026

Uh oh!

chris-maes Mar 16, 2026

Uh oh!

chris-maes left a comment

Uh oh!

anandhkb commented Mar 16, 2026

Uh oh!

anandhkb commented Mar 16, 2026

Uh oh!

chris-maes commented Mar 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

anandhkb commented Mar 14, 2026

Uh oh!

copy-pr-bot bot commented Mar 14, 2026

Uh oh!

coderabbitai bot commented Mar 14, 2026

Walkthrough

Changes

Estimated code review effort

❌ Failed checks (1 warning)

Uh oh!

anandhkb commented Mar 14, 2026

Uh oh!

chris-maes Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

anandhkb Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

chris-maes Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

chris-maes left a comment

Choose a reason for hiding this comment

Uh oh!

anandhkb commented Mar 16, 2026

Uh oh!

anandhkb commented Mar 16, 2026

Uh oh!

chris-maes commented Mar 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants