Add root node presolve using Papilo (#234)

This PR implements:
- Adapters to presolve/postolve before/after running the solver
- Fetching of Papilo headers through Cmake
- System install of boost for pip wheel
- Runtime boolean parameter to enable the feature (true by default for MIP, false for LP). 

Notes:
- LP is off by default due to some presolvers not supporting dual post solve. These presolvers can be disabled but reduction quality is impacted.
- Post-solve status can appear as fail due to absolute/relative tolerances. Papilo uses absolute tolerance during post solve checks. Advice is to set relative tol in cuOpt to a very small value.
- TBB is disabled. To be enabled in a follow up PR for parallel presolve.

Fixes #277 
 
Credits to: @rg20

Authors:
  - Hugo Linsenmaier (https://github.com/hlinsen)
  - Rajesh Gandham (https://github.com/rg20)
  - Ramakrishnap (https://github.com/rgsl888prabhu)
  - Trevor McKay (https://github.com/tmckayus)

Approvers:
  - Alice Boucher (https://github.com/aliceb-nv)
  - Ramakrishnap (https://github.com/rgsl888prabhu)
  - Robert Maynard (https://github.com/robertmaynard)

URL: #234

Author: hlinsen

benchmarks/linear_programming/cuopt/run_mip.cpp

@@ -360,7 +360,7 @@ int main(int argc, char* argv[])
 
   std::string out_dir;
   std::string result_file;
-  int batch_num;
+  int batch_num = -1;
 
   bool heuristics_only = program.get<std::string>("--heuristics-only")[0] == 't';
   int num_cpu_threads  = program.get<int>("--num-cpu-threads");
@@ -392,8 +392,7 @@ int main(int argc, char* argv[])
     for (int i = 0; i < n_gpus; ++i) {
       gpu_queue.push(i);
     }
-    int tests_ran          = 0;
-    int n_instances_solved = 0;
+    int tests_ran = 0;
     std::vector<std::string> paths;
     if (run_selected) {
       for (const auto& instance : instances) {
@@ -421,11 +420,11 @@ int main(int argc, char* argv[])
 
     bool static_dispatch = false;
     if (static_dispatch) {
-      for (int i = 0; i < paths.size(); ++i) {
+      for (size_t i = 0; i < paths.size(); ++i) {
         // TODO implement
       }
     } else {
-      for (int i = 0; i < paths.size(); ++i) {
+      for (size_t i = 0; i < paths.size(); ++i) {
         task_queue.push(paths[i]);
       }
       while (!task_queue.empty()) {

benchmarks/linear_programming/run_mps_files.sh

@@ -31,6 +31,7 @@
 #   --mip-heuristics-only : Run mip heuristics only
 #   --write-log-file : Write log file
 #   --num-cpu-threads : Number of CPU threads to use
+#   --presolve : Enable presolve (default: true for MIP problems, false for LP problems)
 #   --batch-num : Batch number.  This allows to split the work across multiple batches uniformly when resources are limited.
 #   --n-batches : Number of batches
 #   --log-to-console : Log to console
@@ -74,6 +75,7 @@ Optional Arguments:
     --mip-heuristics-only  Run mip heuristics only
     --write-log-file   Write log file
     --num-cpu-threads  Number of CPU threads to use
+    --presolve         Enable presolve (default: true for MIP problems, false for LP problems)
     --batch-num        Batch number
     --n-batches        Number of batches
     --log-to-console   Log to console
@@ -107,47 +109,62 @@ fi
 while [[ $# -gt 0 ]]; do
     case $1 in
         --path)
+            echo "MPS_DIR: $2"
             MPS_DIR="$2"
             shift 2
             ;;
         --ngpus)
+            echo "GPU_COUNT: $2"
             GPU_COUNT="$2"
             shift 2
             ;;
         --time-limit)
+            echo "TIME_LIMIT: $2"
             TIME_LIMIT="$2"
             shift 2
             ;;
         --output-dir)
+            echo "OUTPUT_DIR: $2"
             OUTPUT_DIR="$2"
             shift 2
             ;;
         --relaxation)
-            echo "Running relaxation"
+            echo "RELAXATION: true"
             RELAXATION=true
             shift
             ;;
         --mip-heuristics-only)
+            echo "MIP_HEURISTICS_ONLY: true"
             MIP_HEURISTICS_ONLY=true
             shift
             ;;
         --write-log-file)
-            WRITE_LOG_FILE="$2"
-            shift 2
+            echo "WRITE_LOG_FILE: true"
+            WRITE_LOG_FILE=true
+            shift
             ;;
         --num-cpu-threads)
+            echo "NUM_CPU_THREADS: $2"
             NUM_CPU_THREADS="$2"
             shift 2
             ;;
+        --presolve)
+            echo "PRESOLVE: $2"
+            PRESOLVE="$2"
+            shift 2
+            ;;
         --batch-num)
+            echo "BATCH_NUM: $2"
             BATCH_NUM="$2"
             shift 2
             ;;
         --n-batches)
+            echo "N_BATCHES: $2"
             N_BATCHES="$2"
             shift 2
             ;;
         --log-to-console)
+            echo "LOG_TO_CONSOLE: $2"
             LOG_TO_CONSOLE="$2"
             shift 2
             ;;
@@ -173,6 +190,7 @@ RELAXATION=${RELAXATION:-false}
 MIP_HEURISTICS_ONLY=${MIP_HEURISTICS_ONLY:-false}
 WRITE_LOG_FILE=${WRITE_LOG_FILE:-false}
 NUM_CPU_THREADS=${NUM_CPU_THREADS:-1}
+PRESOLVE=${PRESOLVE:-true}
 BATCH_NUM=${BATCH_NUM:-0}
 N_BATCHES=${N_BATCHES:-1}
 LOG_TO_CONSOLE=${LOG_TO_CONSOLE:-true}
@@ -261,9 +279,8 @@ worker() {
         if [ "$RELAXATION" = true ]; then
             args="$args --relaxation"
         fi
-        if [ "$LOG_TO_CONSOLE" = true ]; then
-            args="$args --log-to-console $LOG_TO_CONSOLE"
-        fi
+        args="$args --log-to-console $LOG_TO_CONSOLE"
+        args="$args --presolve $PRESOLVE"
 
         CUDA_VISIBLE_DEVICES=$gpu_id cuopt_cli "$mps_file" --time-limit $TIME_LIMIT $args
     done
@@ -277,4 +294,4 @@ done
 wait
 
 # Remove the index file
-rm -f "$INDEX_FILE"
+rm -f "$INDEX_FILE"

ci/build_wheel_libcuopt.sh

@@ -21,6 +21,9 @@ source rapids-init-pip
 package_name="libcuopt"
 package_dir="python/libcuopt"
 
+# Install Boost
+bash ci/utils/install_boost.sh
+
 export SKBUILD_CMAKE_ARGS="-DCUOPT_BUILD_WHEELS=ON;-DDISABLE_DEPRECATION_WARNING=ON"
 
 # For pull requests we are enabling assert mode.

ci/utils/install_boost.sh

@@ -0,0 +1,40 @@
+#!/bin/bash
+
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+set -euo pipefail
+
+# Install Boost
+if [ -f /etc/os-release ]; then
+    . /etc/os-release
+    if [[ "$ID" == "rocky" ]]; then
+        echo "Detected Rocky Linux. Installing Boost via dnf..."
+        dnf install -y boost-devel
+        if [[ "$(uname -m)" == "x86_64" ]]; then
+            dnf install -y gcc-toolset-14-libquadmath-devel
+        fi
+    elif [[ "$ID" == "ubuntu" ]]; then
+        echo "Detected Ubuntu. Installing Boost via apt..."
+        apt-get update
+        apt-get install -y libboost-dev
+    else
+        echo "Unknown OS: $ID. Please install Boost development libraries manually."
+        exit 1
+    fi
+else
+    echo "/etc/os-release not found. Cannot determine OS. Please install Boost development libraries manually."
+    exit 1
+fi

conda/environments/all_cuda-128_arch-aarch64.yaml

@@ -5,6 +5,7 @@ channels:
 - rapidsai-nightly
 - conda-forge
 dependencies:
+- boost
 - breathe
 - c-compiler
 - ccache

conda/environments/all_cuda-128_arch-x86_64.yaml

@@ -5,6 +5,7 @@ channels:
 - rapidsai-nightly
 - conda-forge
 dependencies:
+- boost
 - breathe
 - c-compiler
 - ccache

conda/recipes/libcuopt/recipe.yaml

@@ -49,6 +49,7 @@ cache:
 
   requirements:
     build:
+      - boost
       - ${{ compiler("c") }}
       - ${{ compiler("cxx") }}
       - ${{ compiler("cuda") }} =${{ cuda_version }}
@@ -68,6 +69,7 @@ cache:
       - libcurand-dev
       - libcusparse-dev
       - cuda-cudart-dev
+      - boost
 
 outputs:
   - package:
@@ -126,6 +128,7 @@ outputs:
         - ${{ stdlib("c") }}
       host:
         - ${{ pin_subpackage("libmps-parser", exact=True) }}
+        - boost
         - cuda-version =${{ cuda_version }}
         - rapids-logger =0.1
         - librmm =${{ dep_minor_version }}
@@ -135,6 +138,7 @@ outputs:
       run:
         - ${{ pin_compatible("cuda-version", upper_bound="x", lower_bound="x") }}
         - ${{ pin_subpackage("libmps-parser", exact=True) }}
+        - boost
         - librmm =${{ dep_minor_version }}
         - cuda-nvrtc
       ignore_run_exports:
@@ -146,6 +150,8 @@ outputs:
           - libcurand
           - libcusparse
           - librmm
+          - libboost
+          - libboost_iostreams
     tests:
     - package_contents:
         files:
@@ -173,6 +179,7 @@ outputs:
       host:
         - ${{ pin_subpackage("libcuopt", exact=True) }}
         - ${{ pin_subpackage("libmps-parser", exact=True) }}
+        - boost
         - gmock ${{ gtest_version }}
         - gtest ${{ gtest_version }}
         - cuda-cudart-dev
@@ -192,6 +199,8 @@ outputs:
           - libcurand
           - libcusparse
           - librmm
+          - libboost
+          - libboost_iostreams
     about:
       homepage: ${{ load_from_file("python/cuopt/pyproject.toml").project.urls.Homepage }}
       license: ${{ load_from_file("python/cuopt/pyproject.toml").project.license.text }}

cpp/CMakeLists.txt

@@ -13,7 +13,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-cmake_minimum_required(VERSION 3.26.4 FATAL_ERROR)
+cmake_minimum_required(VERSION 3.30.4 FATAL_ERROR)
+
+# Add our custom Find modules to the module path
+list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/thirdparty")
 
 include(GNUInstallDirs)
 
@@ -157,6 +160,21 @@ else()
   find_package(RMM REQUIRED)
   find_package(RAFT REQUIRED)
 endif()
+
+FetchContent_Declare(
+  papilo
+  GIT_REPOSITORY "https://github.com/scipopt/papilo.git"
+  GIT_TAG "v2.4.3"
+  SYSTEM
+)
+
+set(BUILD_TESTING OFF CACHE BOOL "Disable test build for papilo")
+set(TBB OFF CACHE BOOL "Disable TBB")
+set(PAPILO_NO_BINARIES ON)
+option(LUSOL "Disable LUSOL" OFF)
+
+FetchContent_MakeAvailable(papilo)
+
 include(${rapids-cmake-dir}/cpm/rapids_logger.cmake)
 # generate logging macros
 rapids_cpm_rapids_logger(BUILD_EXPORT_SET cuopt-exports INSTALL_EXPORT_SET cuopt-exports)
@@ -204,6 +222,9 @@ target_link_options(cuopt PRIVATE "${CUOPT_BINARY_DIR}/fatbin.ld")
 add_library(cuopt::cuopt ALIAS cuopt)
 # ##################################################################################################
 # - include paths ---------------------------------------------------------------------------------
+
+target_include_directories(cuopt SYSTEM PRIVATE "${papilo_SOURCE_DIR}/src" "${papilo_BINARY_DIR}")
+
 target_include_directories(cuopt
   PRIVATE
   "${CMAKE_CURRENT_SOURCE_DIR}/../thirdparty"
@@ -212,6 +233,7 @@ target_include_directories(cuopt
   "$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>"
   "$<BUILD_INTERFACE:${CMAKE_CURRENT_BINARY_DIR}/include>"
   "$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/libmps_parser/include>"
+  INTERFACE
   "$<INSTALL_INTERFACE:include>"
 )
 
@@ -242,6 +264,7 @@ target_link_libraries(cuopt
   ${CUOPT_PRIVATE_CUDA_LIBS}
   )
 
+
 # ##################################################################################################
 # - generate tests --------------------------------------------------------------------------------
 if(BUILD_TESTS)
@@ -357,6 +380,8 @@ target_link_libraries(cuopt_cli
   PUBLIC
   cuopt
   OpenMP::OpenMP_CXX
+  PRIVATE
+  papilo-core
 )
 set_property(TARGET cuopt_cli PROPERTY INSTALL_RPATH "$ORIGIN/../${lib_dir}")
 
@@ -377,6 +402,8 @@ if(BUILD_BENCHMARKS)
     PUBLIC
     cuopt
     OpenMP::OpenMP_CXX
+    PRIVATE
+    papilo-core
   )
 endif()
 

cpp/cuopt_cli.cpp

@@ -122,9 +122,6 @@ int run_single_file(const std::string& file_path,
     (op_problem.get_problem_category() == cuopt::linear_programming::problem_category_t::MIP ||
      op_problem.get_problem_category() == cuopt::linear_programming::problem_category_t::IP);
 
-  bool sol_found = false;
-  double obj_val = std::numeric_limits<double>::infinity();
-
   auto initial_solution =
     initial_solution_file.empty()
       ? std::vector<double>()
@@ -197,6 +194,11 @@ int main(int argc, char* argv[])
     .default_value(false)
     .implicit_value(true);
 
+  program.add_argument("--presolve")
+    .help("enable/disable presolve (default: true for MIP problems, false for LP problems)")
+    .default_value(true)
+    .implicit_value(true);
+
   std::map<std::string, std::string> arg_name_to_param_name;
   {
     // Add all solver settings as arguments

cpp/include/cuopt/linear_programming/constants.h

@@ -50,13 +50,15 @@
 #define CUOPT_LOG_FILE                    "log_file"
 #define CUOPT_LOG_TO_CONSOLE              "log_to_console"
 #define CUOPT_CROSSOVER                   "crossover"
+#define CUOPT_PRESOLVE                    "presolve"
 #define CUOPT_MIP_ABSOLUTE_TOLERANCE      "mip_absolute_tolerance"
 #define CUOPT_MIP_RELATIVE_TOLERANCE      "mip_relative_tolerance"
 #define CUOPT_MIP_INTEGRALITY_TOLERANCE   "mip_integrality_tolerance"
 #define CUOPT_MIP_ABSOLUTE_GAP            "mip_absolute_gap"
 #define CUOPT_MIP_RELATIVE_GAP            "mip_relative_gap"
 #define CUOPT_MIP_HEURISTICS_ONLY         "mip_heuristics_only"
 #define CUOPT_MIP_SCALING                 "mip_scaling"
+#define CUOPT_MIP_PRESOLVE                "mip_presolve"
 #define CUOPT_SOLUTION_FILE               "solution_file"
 #define CUOPT_NUM_CPU_THREADS             "num_cpu_threads"
 #define CUOPT_USER_PROBLEM_FILE           "user_problem_file"