Recursive RINS (#1482)

This PR implements the recursive RINS heuristics. More specifically, it applies a dive-like procedure to create a RINS sub-MIP, presolved it with Papilo and then passes to the cut-and-branch procedure. It also launch a CPU FJ worker over the sub-MIP that runs in parallel to cut-and-branch.

Since the current implementation, does not have an unified problem representation, this PR also adds:
- A routine for converting a `lp_problem_t` (standard form) to `user_problem_t` (range form).
- Allow Papilo to be applied to an `user_problem_t` instead of `optimization_problem_t`.
- Adds an upper bound callback for retrieving the upper bound of the main solver (since it may change if another worker find a better incumbent).

Additionally, the number of worker per diving heuristic is no longer tied to the total number of workers, which allows the solver to use all diving heuristics in low thread counts (we rotate between the different heuristics for each diving worker launch).


## Results
MIPLIB2017, 10min, GH200

```
================================================================================
 main-2026-07-03 (1) vs recursive-submip (2)
================================================================================

------------------------------------------------------------------------------------------------------------------------------
|                                        |       Run 1        |       Run 2        |     Abs. Diff.     |   Rel. Diff. (%)   |
------------------------------------------------------------------------------------------------------------------------------
| Imported                                                 240                  240                   +0                 --- |
| Feasible                                                 227                  228                   +1                 --- |
| Optimal                                                   86                   91                   +5                 --- |
| Solutions with <0.1% primal gap                          138                  151                  +13                 --- |
| Nodes explored (mean)                              1.326e+07            1.296e+07           -3.057e+05               -2.31 |
| Nodes explored (shifted geomean)                   1.262e+04            1.239e+04                 -230               -1.82 |
| Relative MIP gap (mean)                               0.2777               0.2726             -0.00512               -1.84 |
| Relative MIP gap (shifted geomean)                   0.09089              0.08901            -0.001883               -2.07 |
| Solve time (mean)                                      422.6                435.4               +12.79               +3.03 |
| Solve time (shifted geomean)                           196.9                226.6               +29.66               +15.1 |
| Primal gap (mean)                                      9.969                 9.29               -0.679               -6.81 |
| Primal gap (shifted geomean)                          0.4584               0.3491              -0.1093               -23.8 |
| Primal integral (mean)                                 25.02                22.97               -2.051                -8.2 |
| Primal integral (shifted geomean)                      2.849                2.421              -0.4275                 -15 |
------------------------------------------------------------------------------------------------------------------------------


----------------------------------------------------------------------
|             Name             |     status 1     |     status 2     |
----------------------------------------------------------------------
| neos-1171737                           feasible            optimal |
| radiationm40-10-02                     feasible            optimal |
| rail01                                  timeout           feasible |
| ran14x18-disj-8                        feasible            optimal |
| triptim1                               feasible            optimal |
| unitcal_7                              feasible            optimal |
----------------------------------------------------------------------
```

Authors:
  - Nicolas L. Guidotti (https://github.com/nguidotti)

Approvers:
  - Alice Boucher (https://github.com/aliceb-nv)
  - Chris Maes (https://github.com/chris-maes)
  - Ramakrishna Prabhu (https://github.com/ramakrishnap-nv)

URL: #1482

Author: nguidotti

cpp/CMakeLists.txt

@@ -252,14 +252,14 @@ endif ()
 
 FetchContent_Declare(
         papilo
-        GIT_REPOSITORY "https://github.com/scipopt/papilo.git"
+        GIT_REPOSITORY "https://github.com/nguidotti/papilo.git"
         # We would want to get the main branch. However, the main branch
         # does not have some of the presolvers and settings that we need
         # Mainly, probing and clique merging.
         # This is the reason we are using the development branch
         # from Oct 12, 2025. Once these changes are merged into the main branch,
         #we can switch to the main branch.
-        GIT_TAG "741a2b9c8155b249d6df574d758b4d97d4417520"
+        GIT_TAG "0cfea20c5655249174dbd04f0fb7bd6b1f6e9a0c"
         GIT_PROGRESS TRUE
         EXCLUDE_FROM_ALL
         SYSTEM
@@ -293,13 +293,13 @@ set(BUILD_SHARED_LIBS ${BUILD_SHARED_LIBS_SAVED})
 # dejavu - header-only graph automorphism library for MIP symmetry detection
 # https://github.com/markusa4/dejavu (header-only, skip its CMakeLists.txt)
 FetchContent_Declare(
-  dejavu
-  GIT_REPOSITORY "https://github.com/markusa4/dejavu.git"
-  GIT_TAG "v2.1"
-  GIT_PROGRESS TRUE
-  EXCLUDE_FROM_ALL
-  SYSTEM
-  SOURCE_SUBDIR _nonexistent
+        dejavu
+        GIT_REPOSITORY "https://github.com/markusa4/dejavu.git"
+        GIT_TAG "v2.1"
+        GIT_PROGRESS TRUE
+        EXCLUDE_FROM_ALL
+        SYSTEM
+        SOURCE_SUBDIR _nonexistent
 )
 FetchContent_MakeAvailable(dejavu)
 message(STATUS "dejavu (graph automorphism): ${dejavu_SOURCE_DIR}")
@@ -369,7 +369,7 @@ if (NOT SKIP_GRPC_BUILD)
 
     # Proto search paths: manual protos in src/grpc, generated data proto in src/grpc/codegen/generated
     set(PROTO_PATH_MANUAL "${CMAKE_CURRENT_SOURCE_DIR}/src/grpc")
-    set(PROTO_PATH_GEN    "${CMAKE_CURRENT_SOURCE_DIR}/src/grpc/codegen/generated")
+    set(PROTO_PATH_GEN "${CMAKE_CURRENT_SOURCE_DIR}/src/grpc/codegen/generated")
 
     # Generate C++ code from cuopt_remote_data.proto (auto-generated data definitions)
     set(DATA_PROTO_FILE "${PROTO_PATH_GEN}/cuopt_remote_data.proto")
@@ -441,8 +441,8 @@ add_subdirectory(src)
 # nvcc 13.0.3 ICE (signal 11) compiling sliding_window.cu with 7 GPU architectures;
 # --split-compile breaks the codegen into per-arch sub-jobs to avoid the crash
 set_source_files_properties(
-    ${CMAKE_CURRENT_SOURCE_DIR}/src/routing/local_search/sliding_window.cu
-    PROPERTIES COMPILE_OPTIONS "--split-compile=0")
+        ${CMAKE_CURRENT_SOURCE_DIR}/src/routing/local_search/sliding_window.cu
+        PROPERTIES COMPILE_OPTIONS "--split-compile=0")
 
 if (HOST_LINEINFO)
     set_source_files_properties(${CUOPT_SRC_FILES} DIRECTORY ${CMAKE_SOURCE_DIR} PROPERTIES COMPILE_OPTIONS "-g1")
@@ -495,8 +495,8 @@ set_target_properties(cuopt
 )
 
 target_compile_definitions(cuopt
-  PUBLIC "CUOPT_LOG_ACTIVE_LEVEL=RAPIDS_LOGGER_LOG_LEVEL_${LIBCUOPT_LOGGING_LEVEL}"
-  PUBLIC CUSPARSE_ENABLE_EXPERIMENTAL_API
+        PUBLIC "CUOPT_LOG_ACTIVE_LEVEL=RAPIDS_LOGGER_LOG_LEVEL_${LIBCUOPT_LOGGING_LEVEL}"
+        PUBLIC CUSPARSE_ENABLE_EXPERIMENTAL_API
 )
 
 target_compile_options(cuopt
@@ -528,8 +528,8 @@ target_include_directories(cuopt PRIVATE
 )
 
 target_include_directories(cuopt SYSTEM PRIVATE
-  "${pslp_SOURCE_DIR}/include"
-  "${dejavu_SOURCE_DIR}"
+        "${pslp_SOURCE_DIR}/include"
+        "${dejavu_SOURCE_DIR}"
 )
 
 target_include_directories(cuopt

cpp/include/cuopt/mathematical_optimization/constants.h

@@ -75,6 +75,7 @@
 #define CUOPT_MIP_ZERO_HALF_CUTS                   "mip_zero_half_cuts"
 #define CUOPT_MIP_STRONG_CHVATAL_GOMORY_CUTS       "mip_strong_chvatal_gomory_cuts"
 #define CUOPT_MIP_REDUCED_COST_STRENGTHENING       "mip_reduced_cost_strengthening"
+#define CUOPT_MIP_RINS                             "mip_rins"
 #define CUOPT_MIP_OBJECTIVE_STEP                   "mip_objective_step"
 #define CUOPT_MIP_CUT_CHANGE_THRESHOLD             "mip_cut_change_threshold"
 #define CUOPT_MIP_CUT_MIN_ORTHOGONALITY            "mip_cut_min_orthogonality"
@@ -127,9 +128,19 @@
 #define CUOPT_MIP_HYPER_DIVING_NODE_LIMIT             "mip_hyper_diving_node_limit"
 #define CUOPT_MIP_HYPER_DIVING_ITERATION_LIMIT_FACTOR "mip_hyper_diving_iteration_limit_factor"
 #define CUOPT_MIP_HYPER_DIVING_BACKTRACK_LIMIT        "mip_hyper_diving_backtrack_limit"
-/* @brief Show per-strategy diving symbol in logs (true) instead of a generic 'D' */
+/* @brief Show per-strategy diving symbol in logs instead of a generic 'D' */
 #define CUOPT_MIP_HYPER_DIVING_SHOW_TYPE "mip_hyper_diving_show_type"
 
+/* @brief Recursive sub-MIP (RINS) hyper-parameters */
+#define CUOPT_MIP_HYPER_SUBMIP_BASE_TARGET_FIXRATE   "mip_hyper_submip_base_target_fixrate"
+#define CUOPT_MIP_HYPER_SUBMIP_MIN_FIXRATE           "mip_hyper_submip_min_fixrate"
+#define CUOPT_MIP_HYPER_SUBMIP_MIN_FIXRATE_CAP       "mip_hyper_submip_min_fixrate_cap"
+#define CUOPT_MIP_HYPER_SUBMIP_TARGET_MIP_GAP        "mip_hyper_submip_target_mip_gap"
+#define CUOPT_MIP_HYPER_SUBMIP_NODE_LIMIT_BASE       "mip_hyper_submip_node_limit_base"
+#define CUOPT_MIP_HYPER_SUBMIP_MAX_LEVEL             "mip_hyper_submip_max_level"
+#define CUOPT_MIP_HYPER_SUBMIP_ITERATION_LIMIT_RATIO "mip_hyper_submip_iteration_limit_ratio"
+#define CUOPT_MIP_HYPER_SUBMIP_ENABLE_CPUFJ          "mip_hyper_submip_enable_cpufj"
+
 /* @brief MIP determinism mode constants */
 #define CUOPT_MODE_OPPORTUNISTIC 0
 #define CUOPT_MODE_DETERMINISTIC 1

cpp/include/cuopt/mathematical_optimization/mip/diving_hyper_params.hpp

@@ -10,12 +10,7 @@
 namespace cuopt::mathematical_optimization {
 
 /**
- * @brief Tuning knobs for the dual-simplex diving heuristics used in MIP B&B.
- *
- * Used directly as simplex_solver_settings_t::diving_settings and copied into
- * branch_and_bound_settings.diving_settings before solve. These are registered
- * in the unified parameter framework via solver_settings_t and can be loaded
- * from a config file with load_parameters_from_file().
+ * @brief Tuning knobs for the diving heuristics for the MIP solver.
  */
 template <typename i_t, typename f_t>
 struct mip_diving_hyper_params_t {

cpp/include/cuopt/mathematical_optimization/mip/solver_settings.hpp

@@ -13,6 +13,7 @@
 #include <cuopt/mathematical_optimization/constants.h>
 #include <cuopt/mathematical_optimization/mip/diving_hyper_params.hpp>
 #include <cuopt/mathematical_optimization/mip/heuristics_hyper_params.hpp>
+#include <cuopt/mathematical_optimization/mip/submip_hyper_params.hpp>
 #include <cuopt/mathematical_optimization/pdlp/pdlp_hyper_params.cuh>
 #include <cuopt/mathematical_optimization/utilities/internals.hpp>
 
@@ -184,8 +185,8 @@ class mip_solver_settings_t {
   pdlp::pdlp_hyper_params_t hyper_params;
 
   mip_heuristics_hyper_params_t<i_t, f_t> heuristic_params;
-
   mip_diving_hyper_params_t<i_t, f_t> diving_params;
+  mip_submip_hyper_params_t<i_t, f_t> submip_params;
 
  private:
   std::vector<internals::base_solution_callback_t*> mip_callbacks_;

cpp/include/cuopt/mathematical_optimization/mip/submip_hyper_params.hpp

@@ -0,0 +1,46 @@
+/* clang-format off */
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+/* clang-format on */
+
+#pragma once
+
+/**
+ * @brief Tuning knobs for the recursive sub-MIP.
+ */
+template <typename i_t, typename f_t>
+struct mip_submip_hyper_params_t {
+  // Enable or disable (recursive) RINS: -1 automatic, 0 disabled, 1 enabled
+  i_t rins = -1;
+
+  // Base for calculating the target fix rate for the neighbourhood. Actual target value is
+  // determined automatically according to the success and infeasible rate.
+  f_t base_target_fixrate = 0.6;
+
+  // Minimum fix rate for accepting the neighbourhood.
+  f_t min_fixrate = 0.25;
+
+  // Hard cap for the minimum fix rate for solving a sub-MIP.
+  f_t min_fixrate_cap = 0.1;
+
+  // MIP gap for the sub-MIP (unless the MIP gap from the B&B is lower)
+  f_t target_mip_gap = 0.01;
+
+  // The base node limit for the sub-MIP
+  i_t node_limit_base = 200;
+
+  // The current level in the recursion. This is an internal parameter and will set automatically.
+  i_t level = 0;
+
+  // Maximum recursion level
+  i_t max_level = 10;
+
+  // Limit the number of simplex iterations spent in the submip. Set as a factor of the total
+  // number of simplex iteration from the parent B&B.
+  f_t iteration_limit_ratio = 0.8;
+
+  // Run CPU FJ over the sub-MIP
+  bool enable_cpufj = true;
+};