[DO NOT MERGE] Clique Table and Preprocessing by akifcorduk · Pull Request #627 · NVIDIA/cuopt

akifcorduk · 2025-11-25T15:09:38Z

This PR implements the first part of the paper from the paper: Preprocessing and Cutting Planes with Conflict Graphs.
This part contains only the preprocessing parts and clique cuts will follow in a separate PR:

Clique detection by converting constraints into sorted knapsack constraints. This allows fast clique detection.
Additional cliques in the same constraint by utilizing sorting structures.
Clique extension/merging across the conflict graph.
Clique covering and problem modification.

The data structures and query functions are implemented and will be used as a basis for clique cuts and usage in heuristics.

Benchmark results: there is little to no change in benchmarks. This PR is a basis for the clique cuts PR.

Summary by CodeRabbit

New Features
- Automatic clique detection and injection into MIP presolve to improve preprocessing.
- Host-to-solver constraint sync allowing external problem updates to be applied before solving.
Improvements
- Presolve now respects a time cap to bound preprocessing work.
- Improved constraint handling and propagation during solution setup.
- Local search initialization refined for more reliable first-run behavior.

copy-pr-bot · 2025-11-25T15:09:42Z

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

coderabbitai · 2025-11-25T15:09:59Z

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

@coderabbitai resume to resume automatic reviews.
@coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

▶️ Resume reviews
🔍 Trigger review

📝 Walkthrough

Walkthrough

Adds a clique-based presolve subsystem and integrates it into the MIP pipeline, extends CSR sparse-matrix API, exposes a new problem setter for host-derived constraints, adjusts build files, and updates presolve/solver/local-search flows and comments.

Changes

Cohort / File(s)	Summary
Build Configuration `cpp/CMakeLists.txt`, `cpp/src/mip/CMakeLists.txt`	Updated `solve_MIP` include directories (PRIVATE src, PUBLIC BUILD_INTERFACE include) and added `clique_table.cu` to non-LP build list.
Clique Infrastructure `cpp/src/mip/presolve/conflict_graph/clique_table.cuh`, `cpp/src/mip/presolve/conflict_graph/clique_table.cu`	New templated clique table/types and extensive algorithms for extracting, extending, pruning, and inserting cliques from knapsack-like constraints; explicit instantiations for float/double.
Sparse Matrix API `cpp/src/dual_simplex/sparse_matrix.hpp`, `cpp/src/dual_simplex/sparse_matrix.cpp`	Added `csr_matrix_t::get_constraint_range()` and `insert_row()` overloads; added CSC→CSR scatter helper; fixed comment typo and updated copyright year.
Presolve / Diversity Integration `cpp/src/mip/diversity/diversity_manager.cu`	Hooked initial-clique finding into `run_presolve`: construct host_problem, compute initial cliques, update host constraints via `set_constraints_from_host_user_problem`, reorder/gate bounds updates and trivial presolve.
Probing Cache / Comments `cpp/src/mip/presolve/probing_cache.cu`, `cpp/src/mip/presolve/probing_cache.cuh`	Added `unordered_set` include; added explanatory comments about probing cache behavior (no API/behavioral changes).
Problem Representation `cpp/src/mip/problem/problem.cuh`, `cpp/src/mip/problem/problem.cu`	Added `problem_t::set_constraints_from_host_user_problem(...)` to populate device CSR, rhs, bounds, range handling, names, and variable types from a host user_problem; `resize_constraints` now preserves/initializes prev_dual elements. (Note: duplicate definition observed in cu file.)
Solver Integration `cpp/src/mip/solver.cu`	Introduced presolve time cap (min(10% remaining, 60s)) and call to `context.problem_ptr->set_constraints_from_host_user_problem(branch_and_bound_problem)` before branch-and-bound setup.
Local Search Initialization `cpp/src/mip/local_search/local_search.cu`	On first run, initialize `in_fj.cstr_weights` to constraint count and fill with 1.0.

Estimated code review effort

🎯 5 (Critical) | ⏱️ ~120 minutes

🚥 Pre-merge checks | ✅ 1 | ❌ 2

❌ Failed checks (2 warnings)

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.
Title check	⚠️ Warning	The PR title '[DO NOT MERGE] Clique Table and Preprocessing' uses the '[DO NOT MERGE]' prefix, which is a strong indicator of work-in-progress status and is not appropriate for a production-ready changeset.	Remove '[DO NOT MERGE]' prefix and revise the title to clearly describe the main change, such as 'Add clique table and preprocessing for conflict graph' or similar, once the PR is ready for merge.

✅ Passed checks (1 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.

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

✨ Finishing Touches

📝 Generate docstrings (stacked PR)
📝 Generate docstrings (commit on current branch)

🧪 Generate unit tests (beta)

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

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

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

rg20 · 2026-01-22T14:29:05Z

@akifcorduk is this still relevant?

aliceb-nv

Thanks a lot for your awesome work Akif!! This is going to be super helpful in closing the optimality gap.
Just a few minor comments, mostly relevant for possible follow ups, otherwise LGTM

aliceb-nv · 2026-02-19T10:14:19Z

cpp/src/mip_heuristics/presolve/conflict_graph/clique_table.cu

+std::unordered_set<i_t> clique_table_t<i_t, f_t>::get_adj_set_of_var(i_t var_idx)
+{
+  std::unordered_set<i_t> adj_set;
+  for (const auto& clique_idx : var_clique_map_first[var_idx]) {
+    adj_set.insert(first[clique_idx].begin(), first[clique_idx].end());
+  }
+
+  for (const auto& addtl_clique_idx : var_clique_map_addtl[var_idx]) {
+    adj_set.insert(first[addtl_cliques[addtl_clique_idx].clique_idx].begin(),
+                   first[addtl_cliques[addtl_clique_idx].clique_idx].end());
+  }
+
+  for (const auto& adj_vertex : adj_list_small_cliques[var_idx]) {
+    adj_set.insert(adj_vertex);
+  }
+  // Add the complement of var_idx to the adjacency set
+  i_t complement_idx = (var_idx >= n_variables) ? (var_idx - n_variables) : (var_idx + n_variables);
+  adj_set.insert(complement_idx);
+  adj_set.erase(var_idx);
+  return adj_set;


If i understand correctly, all adjacency sets are of size O(n_vars), which should be small enough in most cases to be a non-factor. Do you think we could use bitsets here instead? It seems to me a lot of clique logic can be translated to bitwise ops

Yes definitely. In cuts PR, computing clique lifting with dynamic programming is done with bitsets. I think we can also convert them. But let's do it in a follow up PR as this PR is quite long.

aliceb-nv · 2026-02-19T10:20:19Z

cpp/src/mip_heuristics/presolve/conflict_graph/clique_table.cuh

+  std::vector<addtl_clique_t<i_t, f_t>> addtl_cliques;
+  // TODO figure out the performance of lookup for the following: unordered_set vs vector
+  // keeps the indices of original(first) cliques that contain variable x
+  std::vector<std::unordered_set<i_t>> var_clique_map_first;


side note: If we ever find that we're bottlenecked on many-cliques instances, I found this paper on sparse bitmaps which looks pretty cool :) https://roaringbitmap.org/

Thank you! There are some heuristic limits to sizes for memory and runtime but we can explore the compressing options :)

aliceb-nv · 2026-02-19T10:35:43Z

cpp/src/mip_heuristics/presolve/conflict_graph/clique_table.cu

+        return true;
+      }
+    }
+  }


Should this be done in the var_idx2 -> var_idx1 direction as well? The additional clique structures aren't symmetrical as I understand

Great catch! Yes additional cliques are not symmetrical. Fixed it.

aliceb-nv · 2026-02-19T10:42:27Z

cpp/src/mip_heuristics/presolve/conflict_graph/clique_table.cu

+    adj_set.insert(first[addtl_cliques[addtl_clique_idx].clique_idx].begin(),
+                   first[addtl_cliques[addtl_clique_idx].clique_idx].end());


This adds the entire base clique. Should this take "start_pos_on_clique" instead to do as "check_adjacency" does? (addtl clique is: vertex_idx + first[clique_idx][start_pos_on_clique:])

Great catch! You are right, we are inserting the whole first clique, we should only insert a representative vertex and part of first clique.

aliceb-nv · 2026-02-19T10:52:12Z

cpp/src/mip_heuristics/presolve/conflict_graph/clique_table.cu

+                   clique_table_t<i_t, f_t>& clique_table,
+                   dual_simplex::user_problem_t<i_t, f_t>& problem,
+                   dual_simplex::csr_matrix_t<i_t, f_t>& A,
+                   cuopt::timer_t& timer)


My intuition is that this could be significantly sped up with some bitwise ops (AND intersections, popcounts). But let's just keep this in mind for possible follow ups if this proves to be a bottleneck :)

aliceb-nv · 2026-02-19T10:58:13Z

cpp/src/mip_heuristics/problem/problem.cu

+void problem_t<i_t, f_t>::set_constraints_from_host_user_problem(
+  const cuopt::linear_programming::dual_simplex::user_problem_t<i_t, f_t>& user_problem)
+{
+  raft::common::nvtx::range fun_scope("set_constraints_from_host_user_problem");
+  cuopt_assert(user_problem.handle_ptr == handle_ptr, "handle mismatch");
+  cuopt_assert(user_problem.num_cols == n_variables, "num cols mismatch");
+  n_constraints = user_problem.num_rows;
+  cuopt_assert(user_problem.rhs.size() == static_cast<size_t>(n_constraints), "rhs size mismatch");
+  cuopt_assert(user_problem.row_sense.size() == static_cast<size_t>(n_constraints),
+               "row sense size mismatch");
+  cuopt_assert(user_problem.range_rows.size() == user_problem.range_value.size(),
+               "range rows/value size mismatch");
+


clique_table.cu can perform variable fixings as well (in remove_dominated_cliques), should the variable bounds be copied as well?

Yes, that's correct. Initially variable fixings were not implemented, now we need to update the variable bounds too. Thanks for catching that!

cpp/src/dual_simplex/sparse_matrix.cpp

akifcorduk added 11 commits October 31, 2025 04:40

add initial files

594ec3f

add some comments and file name changes

f10535f

initial data structures

649062c

find all initial cliques

fe4cc7a

remove small cliques

1d46ec9

renumber cliques on addlt

1474bc5

clique extension is working

b82f63f

Merge branch 'main' of github.com:NVIDIA/cuopt into cliques

1289cbc

add extended cliques into formulation

60edd8e

find constraints to remove

103b4c2

wip

adc9c73

akifcorduk added this to the 26.02 milestone Nov 25, 2025

akifcorduk added non-breaking Introduces a non-breaking change improvement Improves an existing functionality labels Nov 25, 2025

Merge branch 'main' of github.com:NVIDIA/cuopt into cliques

5934c12

rgsl888prabhu changed the base branch from main to release/26.02 January 22, 2026 16:48

akifcorduk added 5 commits January 23, 2026 05:12

Merge branch 'main' of github.com:NVIDIA/cuopt into cliques

43f2ed0

habdle range constraints

3001b52

fix bugs and covert to gpu problem

b45c27e

fix a log

21e565a

move preprocessing to presolve

3ae6047

akifcorduk marked this pull request as ready for review January 29, 2026 11:59

akifcorduk requested review from a team as code owners January 29, 2026 11:59

akifcorduk requested review from Iroy30 and jakirkham January 29, 2026 11:59

aliceb-nv and others added 20 commits February 11, 2026 15:58

disable jobserver unless explicitely requested

ee54477

better workaround fix build

f876fc0

Merge branch 'main' of github.com:NVIDIA/cuopt into fix_timer

1c02baa

add timers to right_looking_lu and refactoring the basis

7de08d2

remove timers from cuts

942de9c

convert lambda to function and remove unnecessary checks

0b944d1

Merge branch 'main' of github.com:NVIDIA/cuopt into cliques

01d68c3

fix merge conflicts, reduce timers

ba1df82

fix reverse_iterator

f395672

fix thrust changes

71b7f2f

handle review comments

82b2d64

add extension heuristics and fix mip gap issues

d074884

handle review comments

0c81173

move timer with inout parameters

609c578

fix merge conflicts

3d32acd

fix merge conflicts

0c54ecf

Merge branch 'main' of github.com:NVIDIA/cuopt into cliques

d01fd3b

handle ai reviewS

d5791dc

revert cmake comment

fc414e7

Merge branch 'fix_timer' into cliques

fff860c

aliceb-nv approved these changes Feb 19, 2026

View reviewed changes

akifcorduk added 5 commits February 20, 2026 02:29

Merge branch 'main' of github.com:NVIDIA/cuopt into cliques

9803df2

fix adjacency checks

3acb6a9

fix adjacency set and var degree

6f6783b

add copy of variable bounds

8bc9fb0

remove the set packing if it covers set partitioning

ba7b9ff

rgsl888prabhu approved these changes Feb 20, 2026

View reviewed changes

chris-maes reviewed Feb 20, 2026

View reviewed changes

cpp/src/dual_simplex/sparse_matrix.cpp Outdated Show resolved Hide resolved

use append row

d82e592

akifcorduk changed the title ~~Clique Table and Preprocessing~~ [DO NOT MERGE] Clique Table and Preprocessing Feb 23, 2026

		adj_set.insert(first[addtl_cliques[addtl_clique_idx].clique_idx].begin(),
		first[addtl_cliques[addtl_clique_idx].clique_idx].end());

Comments

Conversation

akifcorduk commented Nov 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary by CodeRabbit

Uh oh!

copy-pr-bot bot commented Nov 25, 2025

Uh oh!

coderabbitai bot commented Nov 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Reviews paused

Walkthrough

Changes

Estimated code review effort

❌ Failed checks (2 warnings)

Uh oh!

rg20 commented Jan 22, 2026

Uh oh!

aliceb-nv 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!

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!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

akifcorduk commented Nov 25, 2025 •

edited

Loading

coderabbitai bot commented Nov 25, 2025 •

edited

Loading