Add solver caching to support re-solves for barrier QP

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

@@ -22,9 +22,12 @@
 
 #include <cuda/std/span>
 
+
 namespace cuopt {
 namespace CUOPT_EXPORT mathematical_optimization {
 
+class barrier_cache_t;
+
 // Forward declare solver_settings_t for friend class
 template <typename i_t, typename f_t>
 class solver_settings_t;
@@ -365,6 +368,10 @@ class pdlp_solver_settings_t {
   // Used to force batch PDLP to solve a subbatch of the problems at a time
   // The 0 default value will make the solver use its heuristic to determine the subbatch size
   i_t fixed_batch_size{0};
+  /** When true, first GPU barrier/QCQP solve returns a ``barrier_cache_t`` capsule. */
+  bool sequence_solve{false};
+  /** Non-owning cache pointer set by ``call_solve`` for barrier symbolic reuse. */
+  barrier_cache_t* barrier_cache{nullptr};
 
  private:
   /** Initial primal solution */

cpp/include/cuopt/mathematical_optimization/utilities/barrier_cache.hpp

@@ -0,0 +1,84 @@
+/* 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
+
+#include <cuopt/export.hpp>
+
+#include <memory>
+
+#include <raft/core/handle.hpp>
+#include <rmm/cuda_stream.hpp>
+
+namespace cuopt::mathematical_optimization::barrier {
+template <typename i_t, typename f_t>
+class iteration_data_t;
+
+void destroy_iteration_data(iteration_data_t<int, double>* data);
+
+void apply_barrier_linear_objective(iteration_data_t<int, double>& data,
+                                    double const* barrier_c,
+                                    int n);
+}  // namespace cuopt::mathematical_optimization::barrier
+
+namespace cuopt {
+namespace CUOPT_EXPORT mathematical_optimization {
+
+struct barrier_transform_t;
+
+/**
+ * @brief GPU solve cache owned by DataModel when sequence_solve is on.
+ *
+ * After an Optimal full solve, holds iteration_data_t and the user-barrier transform.
+ * update_linear_objective crushes the new linear objective and sets c_dirty so the next Solve
+ * reuses that workspace (skip convert/presolve/scaling).
+ */
+class barrier_cache_t {
+ public:
+  static std::unique_ptr<barrier_cache_t> create(unsigned stream_flags);
+
+  barrier_cache_t(barrier_cache_t&&) noexcept;
+  barrier_cache_t& operator=(barrier_cache_t&&) noexcept;
+  ~barrier_cache_t();
+
+  [[nodiscard]] raft::handle_t* handle_ptr();
+  [[nodiscard]] raft::handle_t const* handle_ptr() const;
+
+  /** Drop cached iteration workspace and transform (handle/stream stay). */
+  void clear();
+
+  /**
+   * @brief Take ownership of barrier iteration workspace. @p data may be null (clears).
+   */
+  void store_iteration_data(barrier::iteration_data_t<int, double>* data);
+
+  /**
+   * @brief Release ownership of cached iteration workspace; caller must delete or wrap it.
+   */
+  barrier::iteration_data_t<int, double>* release_iteration_data();
+
+  void store_transform(std::unique_ptr<barrier_transform_t> transform);
+  [[nodiscard]] barrier_transform_t* transform();
+  [[nodiscard]] barrier_transform_t const* transform() const;
+  void set_c_dirty(bool dirty);
+  [[nodiscard]] bool c_dirty() const;
+
+  /**
+   * Crush the input linear objective into cached iteration_data_t.c / d_c_ and set c_dirty.
+   * Requires a stored transform and iteration_data from an Optimal solve.
+   */
+  void update_linear_objective(double const* c, int n);
+
+ private:
+  barrier_cache_t(std::unique_ptr<rmm::cuda_stream> stream, std::unique_ptr<raft::handle_t> handle);
+
+  struct impl;
+  std::unique_ptr<impl> impl_;
+};
+
+}  // namespace CUOPT_EXPORT mathematical_optimization
+}  // namespace cuopt

cpp/include/cuopt/mathematical_optimization/utilities/cython_solve.hpp

@@ -12,7 +12,6 @@
 #include <cuopt/mathematical_optimization/optimization_problem_solution_interface.hpp>
 #include <cuopt/mathematical_optimization/solver_settings.hpp>
 #include <cuopt/mathematical_optimization/utilities/cython_types.hpp>
-
 #include <cuopt/mathematical_optimization/io/data_model_view.hpp>
 #include <memory>
 #include <raft/core/handle.hpp>
@@ -56,7 +55,8 @@ std::unique_ptr<solver_ret_t> call_solve(
   cuopt::mathematical_optimization::io::data_model_view_t<int, double>*,
   mathematical_optimization::solver_settings_t<int, double>*,
   unsigned int flags = cudaStreamNonBlocking,
-  bool is_batch_mode = false);
+  bool is_batch_mode = false,
+  mathematical_optimization::barrier_cache_t* cache_in = nullptr);
 
 std::pair<std::vector<std::unique_ptr<solver_ret_t>>, double> solve_batch_remote(
   std::vector<cuopt::mathematical_optimization::io::data_model_view_t<int, double>*>,

cpp/include/cuopt/mathematical_optimization/utilities/cython_types.hpp

@@ -10,6 +10,7 @@
 #include <cuopt/export.hpp>
 #include <cuopt/mathematical_optimization/mip/solver_solution.hpp>
 #include <cuopt/mathematical_optimization/pdlp/solver_solution.hpp>
+#include <cuopt/mathematical_optimization/utilities/barrier_cache.hpp>
 #include <cuopt/mathematical_optimization/utilities/internals.hpp>
 
 #include <rmm/device_buffer.hpp>
@@ -86,6 +87,10 @@ struct linear_programming_ret_t {
   double solve_time_{};
   mathematical_optimization::method_t solved_by_{};
 
+  /** GPU barrier cache (stream + handle + iteration workspace); moved to Python capsule when set.
+   */
+  std::unique_ptr<mathematical_optimization::barrier_cache_t> barrier_cache;
+
   bool is_gpu() const { return std::holds_alternative<gpu_solutions_t>(solutions_); }
 };
 

cpp/src/barrier/CMakeLists.txt

@@ -6,6 +6,7 @@
 set(BARRIER_SRC_FILES
   ${CMAKE_CURRENT_SOURCE_DIR}/cusparse_view.cu
   ${CMAKE_CURRENT_SOURCE_DIR}/barrier.cu
+  ${CMAKE_CURRENT_SOURCE_DIR}/barrier_cache.cu
   ${CMAKE_CURRENT_SOURCE_DIR}/device_sparse_matrix.cu
   ${CMAKE_CURRENT_SOURCE_DIR}/pinned_host_allocator.cu
   )