Add memory model support for host-resident problem data

.github/release.yml

@@ -0,0 +1,24 @@
+changelog:
+  exclude:
+    labels:
+      - ignore-for-release
+      - dependencies
+    authors:
+      - rapids-bot[bot]
+      - dependabot[bot]
+  categories:
+    - title: "\U0001F6A8 Breaking Changes"
+      labels:
+        - breaking
+    - title: "\U0001F41B Bug Fixes"
+      labels:
+        - bug
+    - title: "\U0001F4D6 Documentation"
+      labels:
+        - doc
+    - title: "\U0001F680 New Features"
+      labels:
+        - feature request
+    - title: "\U0001F6E0\uFE0F Improvements"
+      labels:
+        - improvement

ci/run_ctests.sh

@@ -1,5 +1,5 @@
 #!/bin/bash
-# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
 set -euo pipefail
@@ -26,3 +26,11 @@ for gt in "${GTEST_DIR}"/*_TEST; do
     echo "Running gtest ${test_name}"
     "${gt}" "$@"
 done
+
+# Run C_API_TEST with CPU memory for local solves (excluding time limit tests)
+if [ -x "${GTEST_DIR}/C_API_TEST" ]; then
+  echo "Running gtest C_API_TEST with CUOPT_USE_CPU_MEM_FOR_LOCAL"
+  CUOPT_USE_CPU_MEM_FOR_LOCAL=1 "${GTEST_DIR}/C_API_TEST" --gtest_filter=-c_api/TimeLimitTestFixture.* "$@"
+else
+  echo "Skipping C_API_TEST with CUOPT_USE_CPU_MEM_FOR_LOCAL (binary not found)"
+fi

cpp/cuopt_cli.cpp

@@ -1,12 +1,15 @@
 /* clang-format off */
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  * SPDX-License-Identifier: Apache-2.0
  */
 /* clang-format on */
 
+#include <cuopt/linear_programming/backend_selection.hpp>
 #include <cuopt/linear_programming/mip/solver_settings.hpp>
 #include <cuopt/linear_programming/optimization_problem.hpp>
+#include <cuopt/linear_programming/optimization_problem_interface.hpp>
+#include <cuopt/linear_programming/optimization_problem_utils.hpp>
 #include <cuopt/linear_programming/solve.hpp>
 #include <mps_parser/parser.hpp>
 #include <utilities/logger.hpp>
@@ -89,7 +92,6 @@ int run_single_file(const std::string& file_path,
                     bool solve_relaxation,
                     const std::map<std::string, std::string>& settings_strings)
 {
-  const raft::handle_t handle_{};
   cuopt::linear_programming::solver_settings_t<int, double> settings;
 
   try {
@@ -122,13 +124,31 @@ int run_single_file(const std::string& file_path,
     return -1;
   }
 
-  auto op_problem =
-    cuopt::linear_programming::mps_data_model_to_optimization_problem(&handle_, mps_data_model);
+  // Determine memory backend and create problem using interface
+  // Create handle only for GPU memory backend (avoid CUDA init on CPU-only hosts)
+  auto memory_backend = cuopt::linear_programming::get_memory_backend_type();
+  std::unique_ptr<raft::handle_t> handle_ptr;
+  std::unique_ptr<cuopt::linear_programming::optimization_problem_interface_t<int, double>>
+    problem_interface;
+
+  if (memory_backend == cuopt::linear_programming::memory_backend_t::GPU) {
+    handle_ptr = std::make_unique<raft::handle_t>();
+    problem_interface =
+      std::make_unique<cuopt::linear_programming::gpu_optimization_problem_t<int, double>>(
+        handle_ptr.get());
+  } else {
+    problem_interface =
+      std::make_unique<cuopt::linear_programming::cpu_optimization_problem_t<int, double>>(nullptr);
+  }
+
+  // Populate the problem from MPS data model
+  cuopt::linear_programming::populate_from_mps_data_model(problem_interface.get(), mps_data_model);
 
-  const bool is_mip =
-    (op_problem.get_problem_category() == cuopt::linear_programming::problem_category_t::MIP ||
-     op_problem.get_problem_category() == cuopt::linear_programming::problem_category_t::IP) &&
-    !solve_relaxation;
+  const bool is_mip = (problem_interface->get_problem_category() ==
+                         cuopt::linear_programming::problem_category_t::MIP ||
+                       problem_interface->get_problem_category() ==
+                         cuopt::linear_programming::problem_category_t::IP) &&
+                      !solve_relaxation;
 
   try {
     auto initial_solution =
@@ -157,10 +177,10 @@ int run_single_file(const std::string& file_path,
   try {
     if (is_mip) {
       auto& mip_settings = settings.get_mip_settings();
-      auto solution      = cuopt::linear_programming::solve_mip(op_problem, mip_settings);
+      auto solution = cuopt::linear_programming::solve_mip(problem_interface.get(), mip_settings);
     } else {
       auto& lp_settings = settings.get_pdlp_settings();
-      auto solution     = cuopt::linear_programming::solve_lp(op_problem, lp_settings);
+      auto solution     = cuopt::linear_programming::solve_lp(problem_interface.get(), lp_settings);
     }
   } catch (const std::exception& e) {
     CUOPT_LOG_ERROR("Error: %s", e.what());
@@ -334,19 +354,23 @@ int main(int argc, char* argv[])
   const auto initial_solution_file = program.get<std::string>("--initial-solution");
   const auto solve_relaxation      = program.get<bool>("--relaxation");
 
-  // All arguments are parsed as string, default values are parsed as int if unused.
-  const auto num_gpus = program.is_used("--num-gpus")
-                          ? std::stoi(program.get<std::string>("--num-gpus"))
-                          : program.get<int>("--num-gpus");
-
+  // Only initialize CUDA resources if using GPU memory backend (not remote execution)
+  auto memory_backend = cuopt::linear_programming::get_memory_backend_type();
   std::vector<std::shared_ptr<rmm::mr::device_memory_resource>> memory_resources;
 
-  for (int i = 0; i < std::min(raft::device_setter::get_device_count(), num_gpus); ++i) {
-    cudaSetDevice(i);
-    memory_resources.push_back(make_async());
-    rmm::mr::set_per_device_resource(rmm::cuda_device_id{i}, memory_resources.back().get());
+  if (memory_backend == cuopt::linear_programming::memory_backend_t::GPU) {
+    // All arguments are parsed as string, default values are parsed as int if unused.
+    const auto num_gpus = program.is_used("--num-gpus")
+                            ? std::stoi(program.get<std::string>("--num-gpus"))
+                            : program.get<int>("--num-gpus");
+
+    for (int i = 0; i < std::min(raft::device_setter::get_device_count(), num_gpus); ++i) {
+      RAFT_CUDA_TRY(cudaSetDevice(i));
+      memory_resources.push_back(make_async());
+      rmm::mr::set_per_device_resource(rmm::cuda_device_id{i}, memory_resources.back().get());
+    }
+    RAFT_CUDA_TRY(cudaSetDevice(0));
   }
-  cudaSetDevice(0);
 
   return run_single_file(file_name, initial_solution_file, solve_relaxation, settings_strings);
 }

cpp/include/cuopt/linear_programming/backend_selection.hpp

@@ -0,0 +1,69 @@
+/* clang-format off */
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+/* clang-format on */
+
+#pragma once
+
+namespace cuopt::linear_programming {
+
+/**
+ * @brief Enum for execution mode (local vs remote solve)
+ */
+enum class execution_mode_t {
+  LOCAL,  ///< Solve locally on this machine
+  REMOTE  ///< Solve remotely via gRPC
+};
+
+/**
+ * @brief Enum for memory backend type (GPU vs CPU memory)
+ */
+enum class memory_backend_t {
+  GPU,  ///< Use GPU memory (device memory via RMM)
+  CPU   ///< Use CPU memory (host memory)
+};
+
+/**
+ * @brief Check if remote execution is enabled via environment variables
+ * @return true if both CUOPT_REMOTE_HOST and CUOPT_REMOTE_PORT are set
+ */
+bool is_remote_execution_enabled();
+
+/**
+ * @brief Determine execution mode based on environment variables
+ *
+ * @return execution_mode_t::REMOTE if CUOPT_REMOTE_HOST and CUOPT_REMOTE_PORT are set,
+ *         execution_mode_t::LOCAL otherwise
+ */
+execution_mode_t get_execution_mode();
+
+/**
+ * @brief Check if GPU memory should be used for remote execution
+ * @return true if CUOPT_USE_GPU_MEM_FOR_REMOTE is set to "true" or "1" (case-insensitive)
+ */
+bool use_gpu_memory_for_remote();
+
+/**
+ * @brief Check if CPU memory should be used for local execution (test mode)
+ *
+ * This is intended for testing CPU problem/solution structures without remote execution.
+ * When enabled, local solve will convert CPU problems to GPU, solve, and convert back.
+ *
+ * @return true if CUOPT_USE_CPU_MEM_FOR_LOCAL is set to "true" or "1" (case-insensitive)
+ */
+bool use_cpu_memory_for_local();
+
+/**
+ * @brief Determine which memory backend to use based on execution mode
+ *
+ * Logic:
+ *   - LOCAL execution -> GPU memory by default, CPU if CUOPT_USE_CPU_MEM_FOR_LOCAL=true (test mode)
+ *   - REMOTE execution -> CPU memory by default, GPU if CUOPT_USE_GPU_MEM_FOR_REMOTE=true
+ *
+ * @return memory_backend_t::GPU or memory_backend_t::CPU
+ */
+memory_backend_t get_memory_backend_type();
+
+}  // namespace cuopt::linear_programming