refactor: make solver settings constructible without CUDA

Two things forced CUDA on anything that merely constructed or inspected solver
settings, even when it never touched a device.

1. pdlp_solver_settings_t held pdlp_warm_start_data_t by value. That type owns
   nine rmm::device_uvector, and its default constructor is out-of-line in a CUDA
   translation unit because device_uvector has no default ctor -- it needs a
   stream, and building even a zero-size one calls cudaGetDevice. So constructing
   settings pulled in libcuopt.

   It is now held by shared_ptr, allocated lazily via ensure_pdlp_warm_start_data().
   shared_ptr rather than unique_ptr specifically: shared_ptr type-erases its
   deleter into the control block at construction, so a host-only translation unit
   can copy and destroy the member without the complete type. unique_ptr would
   only move the problem from the constructor to the destructor.

   The ~88 device-side uses inside set_pdlp_warm_start_data() are unchanged; a
   local reference alias keeps that code reading as before.

2. populate_from_data_model_view() inlined both the GPU and CPU warm-start paths
   in one if/else. The GPU direction is only reachable when handle != nullptr, but
   the compiler instantiated both branches into every translation unit including
   the header -- dragging convert_to_gpu_warmstart, convert_to_cpu_warmstart and
   pdlp_warm_start_data_t(view, stream) along with it.

   Split into apply_warmstart_gpu_target() (declared in the header, defined in
   optimization_problem.cu) and apply_warmstart_cpu_target() (host-only, inline),
   selected by a kHostOnly template parameter dispatched with `if constexpr`.
   The compile-time dispatch is the point: a host-only caller never *instantiates*
   the GPU branch, so it emits no reference to it. A runtime `if` would not help.

Also moves the warm-start accessors that need no allocation into
solver_settings_accessors.cpp, leaving the CUDA TU with only members that do.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Signed-off-by: Ramakrishna Prabhu <ramakrishnap@nvidia.com>

Author: ramakrishnap-nv

cpp/include/cuopt/mathematical_optimization/optimization_problem_utils.hpp

@@ -137,6 +137,39 @@ void populate_from_mps_data_model(optimization_problem_interface_t<i_t, f_t>* pr
   }
 }
 
+/**
+ * @brief Move warm-start data into the form a GPU solve needs (H2D / view->device_uvector).
+ *
+ * Declared here, defined in libcuopt (optimization_problem.cu): it touches device memory,
+ * so keeping it out-of-line is what lets CUDA-free consumers of this header link without
+ * a CUDA runtime. Only call it with a real handle.
+ */
+template <typename i_t, typename f_t>
+void apply_warmstart_gpu_target(solver_settings_t<i_t, f_t>* solver_settings,
+                                const raft::handle_t* handle);
+
+/**
+ * @brief Move warm-start data into the form a CPU / remote solve needs.
+ *
+ * Host-only by construction. A CPU-only caller cannot be holding device-resident warm
+ * start (there is no device to have populated it), so that case is rejected rather than
+ * converted -- converting would require a D2H copy and thus CUDA.
+ */
+template <typename i_t, typename f_t>
+void apply_warmstart_cpu_target(solver_settings_t<i_t, f_t>* solver_settings)
+{
+  auto& pdlp = solver_settings->get_pdlp_settings();
+
+  if (pdlp.get_cpu_pdlp_warm_start_data().is_populated()) { return; }
+
+  // Warmstart view (host spans from Cython) -> CPU backend: copy directly, no CUDA needed.
+  if (solver_settings->get_pdlp_warm_start_data_view()
+        .last_restart_duality_gap_dual_solution_.size() > 0) {
+    pdlp.get_cpu_pdlp_warm_start_data() =
+      cpu_pdlp_warm_start_data_t<i_t, f_t>(solver_settings->get_pdlp_warm_start_data_view());
+  }
+}
+
 /**
  * @brief Transfer parsed MPS/QPS storage into a CPU-backed problem without copying payload arrays.
  *
@@ -176,7 +209,7 @@ void adopt_from_mps_data_model(optimization_problem_interface_t<i_t, f_t>* probl
  * @param[in] solver_settings Optional solver settings (for warmstart data, GPU only)
  * @param[in] handle Optional RAFT handle (for warmstart data, GPU only)
  */
-template <typename i_t, typename f_t>
+template <typename i_t, typename f_t, bool kHostOnly = false>
 void populate_from_data_model_view(
   optimization_problem_interface_t<i_t, f_t>* problem,
   cuopt::mathematical_optimization::io::data_model_view_t<i_t, f_t>* data_model,
@@ -209,57 +242,26 @@ void populate_from_data_model_view(
   problem->set_objective_scaling_factor(data_model->get_objective_scaling_factor());
   problem->set_objective_offset(data_model->get_objective_offset());
 
-  // Handle warmstart data with GPU↔CPU conversion if needed
+  // Handle warmstart data with GPU<->CPU conversion if needed.
+  //
+  // Split into two helpers deliberately. The GPU direction is only reachable when
+  // handle != nullptr, but a single inlined if/else instantiated BOTH directions into
+  // every TU that includes this header -- which dragged convert_to_gpu_warmstart,
+  // pdlp_warm_start_data_t(view, stream) and friends into the CUDA-free gRPC client.
+  // apply_warmstart_gpu_target() is declared here and defined in libcuopt, so only
+  // callers that actually pass a handle reference it.
+  //
+  // kHostOnly is a compile-time opt-out, not just a runtime one: `if constexpr` means a
+  // host-only caller never *instantiates* the GPU branch, so it emits no reference to
+  // apply_warmstart_gpu_target and needs no CUDA runtime to link.
   if (solver_settings != nullptr) {
-    bool target_is_gpu = (handle != nullptr);
-
-    // Check which warmstart type is populated
-    // Note: Python sets the VIEW (spans), so check both view and data for GPU warmstart
-    // CPU warmstart is set directly in the data structure
-    bool has_gpu_warmstart_view = (solver_settings->get_pdlp_warm_start_data_view()
-                                     .last_restart_duality_gap_dual_solution_.size() > 0);
-    bool has_gpu_warmstart_data =
-      solver_settings->get_pdlp_settings().get_pdlp_warm_start_data().is_populated();
-    bool has_cpu_warmstart =
-      solver_settings->get_pdlp_settings().get_cpu_pdlp_warm_start_data().is_populated();
-
-    bool has_gpu_warmstart = has_gpu_warmstart_view || has_gpu_warmstart_data;
-
-    if (has_gpu_warmstart || has_cpu_warmstart) {
-      if (target_is_gpu) {
-        // Target is GPU backend
-        if (has_gpu_warmstart_view) {
-          // GPU warmstart from Python → GPU backend: copy view (spans) to data (device_uvectors)
-          // Python sets the view (spans over cuDF), but solver needs device_uvectors
-          pdlp_warm_start_data_t<i_t, f_t> pdlp_warm_start_data(
-            solver_settings->get_pdlp_warm_start_data_view(), handle->get_stream());
-          solver_settings->get_pdlp_settings().set_pdlp_warm_start_data(pdlp_warm_start_data);
-        } else if (has_gpu_warmstart_data) {
-          // GPU warmstart from C++ API → GPU backend: data already set, nothing to do
-          // The device_uvectors are already populated in the settings
-        } else {
-          // CPU warmstart → GPU backend: convert H2D
-          pdlp_warm_start_data_t<i_t, f_t> gpu_warmstart = convert_to_gpu_warmstart(
-            solver_settings->get_pdlp_settings().get_cpu_pdlp_warm_start_data(),
-            handle->get_stream());
-          solver_settings->get_pdlp_settings().set_pdlp_warm_start_data(gpu_warmstart);
-        }
+    if constexpr (kHostOnly) {
+      apply_warmstart_cpu_target(solver_settings);
+    } else {
+      if (handle != nullptr) {
+        apply_warmstart_gpu_target(solver_settings, handle);
       } else {
-        // Target is CPU backend (remote execution)
-        if (has_cpu_warmstart) {
-          // CPU warmstart → CPU backend: data already in correct form, nothing to do
-        } else if (has_gpu_warmstart_view) {
-          // Warmstart view (host spans from Cython) → CPU backend: copy directly, no CUDA needed
-          solver_settings->get_pdlp_settings().get_cpu_pdlp_warm_start_data() =
-            cpu_pdlp_warm_start_data_t<i_t, f_t>(solver_settings->get_pdlp_warm_start_data_view());
-        } else {
-          // GPU warmstart data (device_uvectors) → CPU backend: convert D2H
-          auto& gpu_ws = solver_settings->get_pdlp_settings().get_pdlp_warm_start_data();
-          cpu_pdlp_warm_start_data_t<i_t, f_t> cpu_warmstart =
-            convert_to_cpu_warmstart(gpu_ws, gpu_ws.current_primal_solution_.stream());
-          solver_settings->get_pdlp_settings().get_cpu_pdlp_warm_start_data() =
-            std::move(cpu_warmstart);
-        }
+        apply_warmstart_cpu_target(solver_settings);
       }
     }
   }

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

@@ -13,6 +13,7 @@
 #include <cuopt/mathematical_optimization/pdlp/pdlp_hyper_params.cuh>
 #include <cuopt/mathematical_optimization/pdlp/pdlp_warm_start_data.hpp>
 #include <cuopt/mathematical_optimization/utilities/internals.hpp>
+#include <memory>
 #include <optional>
 #include <raft/core/device_span.hpp>
 #include <rmm/device_uvector.hpp>
@@ -371,8 +372,24 @@ class pdlp_solver_settings_t {
   /** Initial pdlp iteration */
   // TODO batch mode: tmp
   std::optional<i_t> initial_pdlp_iteration_;
-  /** GPU-backed warm start data (device_uvector), used by C++ API and local GPU solves */
-  pdlp_warm_start_data_t<i_t, f_t> pdlp_warm_start_data_;
+  /** GPU-backed warm start data (device_uvector), used by C++ API and local GPU solves.
+   *
+   * Held by shared_ptr rather than by value so that constructing a settings object needs
+   * no CUDA. pdlp_warm_start_data_t owns nine rmm::device_uvector, and its default ctor is
+   * out-of-line in a CUDA TU (device_uvector has no default ctor -- it needs a stream, and
+   * building even a zero-size one calls cudaGetDevice). By value, that made every consumer
+   * of solver_settings_t -- including the CUDA-free gRPC client -- depend on libcuopt.
+   *
+   * shared_ptr specifically, not unique_ptr: shared_ptr type-erases its deleter into the
+   * control block at construction, so a host-only TU can copy and destroy this member
+   * without the complete type. unique_ptr would just move the problem to the destructor.
+   *
+   * Null until a GPU consumer first needs it; use ensure_pdlp_warm_start_data().
+   */
+  mutable std::shared_ptr<pdlp_warm_start_data_t<i_t, f_t>> pdlp_warm_start_data_;
+
+  /** Lazily allocate pdlp_warm_start_data_ and return it. Defined in a CUDA TU. */
+  pdlp_warm_start_data_t<i_t, f_t>& ensure_pdlp_warm_start_data() const;
   /** Warm start data as spans over external memory, used by Cython/Python interface */
   pdlp_warm_start_data_view_t<i_t, f_t> pdlp_warm_start_data_view_;
   /** CPU-backed warm start data (std::vector), used for remote execution on CPU-only hosts */

cpp/src/grpc/client/cython_grpc_client.cpp

@@ -109,7 +109,10 @@ grpc_submit_result_t grpc_python_client_t::submit(
   }
 
   cuopt::mathematical_optimization::cpu_optimization_problem_t<int, double> cpu_problem;
-  cuopt::mathematical_optimization::populate_from_data_model_view(
+  // <int, double, /*kHostOnly=*/true>: this is a remote client, so the GPU warm-start
+  // path is unreachable here. Selecting it explicitly keeps the device conversions from
+  // being instantiated into cuopt_client.
+  cuopt::mathematical_optimization::populate_from_data_model_view<int, double, true>(
     &cpu_problem, data_model, settings, nullptr);
 
   const bool is_mip =

cpp/src/pdlp/CMakeLists.txt

@@ -6,6 +6,7 @@
 # Core LP files always included
 set(LP_CORE_FILES
   ${CMAKE_CURRENT_SOURCE_DIR}/solver_settings.cu
+  ${CMAKE_CURRENT_SOURCE_DIR}/solver_settings_accessors.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/optimization_problem.cu
   ${CMAKE_CURRENT_SOURCE_DIR}/cpu_optimization_problem.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/cpu_optimization_problem_to_gpu.cpp

cpp/src/pdlp/optimization_problem.cu

@@ -1637,4 +1637,43 @@ template CUOPT_EXPORT optimization_problem_t<int32_t, float>
     rmm::cuda_stream_view) const;
 #endif
 
+// GPU-target warm-start handling, declared in optimization_problem_utils.hpp.
+//
+// Defined here rather than inline in the header so that CUDA-free consumers of that
+// header (the gRPC client in cuopt_client) never instantiate the device conversions.
+template <typename i_t, typename f_t>
+void apply_warmstart_gpu_target(solver_settings_t<i_t, f_t>* solver_settings,
+                                const raft::handle_t* handle)
+{
+  auto& pdlp = solver_settings->get_pdlp_settings();
+
+  const bool has_view        = (solver_settings->get_pdlp_warm_start_data_view()
+                           .last_restart_duality_gap_dual_solution_.size() > 0);
+  const bool has_device_data = pdlp.get_pdlp_warm_start_data().is_populated();
+  const bool has_host_data   = pdlp.get_cpu_pdlp_warm_start_data().is_populated();
+
+  if (!has_view && !has_device_data && !has_host_data) { return; }
+
+  if (has_view) {
+    // Warmstart from Python (spans over cuDF) -> solver needs device_uvectors.
+    pdlp_warm_start_data_t<i_t, f_t> warm_start(solver_settings->get_pdlp_warm_start_data_view(),
+                                                handle->get_stream());
+    pdlp.set_pdlp_warm_start_data(warm_start);
+  } else if (has_device_data) {
+    // Already device-resident from the C++ API: nothing to do.
+  } else {
+    // Host warmstart -> GPU backend: convert H2D.
+    pdlp_warm_start_data_t<i_t, f_t> warm_start =
+      convert_to_gpu_warmstart(pdlp.get_cpu_pdlp_warm_start_data(), handle->get_stream());
+    pdlp.set_pdlp_warm_start_data(warm_start);
+  }
+}
+
+#if MIP_INSTANTIATE_FLOAT
+template void apply_warmstart_gpu_target(solver_settings_t<int, float>*, const raft::handle_t*);
+#endif
+#if MIP_INSTANTIATE_DOUBLE
+template void apply_warmstart_gpu_target(solver_settings_t<int, double>*, const raft::handle_t*);
+#endif
+
 }  // namespace cuopt::mathematical_optimization

cpp/src/pdlp/solver_settings.cu

@@ -95,7 +95,10 @@ void pdlp_solver_settings_t<i_t, f_t>::set_pdlp_warm_start_data(
   const rmm::device_uvector<i_t>& var_mapping,
   const rmm::device_uvector<i_t>& constraint_mapping)
 {
-  pdlp_warm_start_data_ = std::move(pdlp_warm_start_data_view);
+  // pdlp_warm_start_data_ is a shared_ptr now (see solver_settings.hpp); alias it so the
+  // device code below reads unchanged.
+  auto& pdlp_warm_start_data_ = ensure_pdlp_warm_start_data();
+  pdlp_warm_start_data_       = std::move(pdlp_warm_start_data_view);
 
   // A var_mapping was given
   if (var_mapping.size() != 0) {
@@ -382,37 +385,28 @@ std::optional<i_t> pdlp_solver_settings_t<i_t, f_t>::get_initial_pdlp_iteration(
 }
 
 template <typename i_t, typename f_t>
-const pdlp_warm_start_data_t<i_t, f_t>& pdlp_solver_settings_t<i_t, f_t>::get_pdlp_warm_start_data()
-  const noexcept
-{
-  return pdlp_warm_start_data_;
-}
-
-template <typename i_t, typename f_t>
-pdlp_warm_start_data_t<i_t, f_t>& pdlp_solver_settings_t<i_t, f_t>::get_pdlp_warm_start_data()
-{
-  return pdlp_warm_start_data_;
-}
-
-template <typename i_t, typename f_t>
-const cpu_pdlp_warm_start_data_t<i_t, f_t>&
-pdlp_solver_settings_t<i_t, f_t>::get_cpu_pdlp_warm_start_data() const noexcept
+pdlp_warm_start_data_t<i_t, f_t>& pdlp_solver_settings_t<i_t, f_t>::ensure_pdlp_warm_start_data()
+  const
 {
-  return cpu_pdlp_warm_start_data_;
+  if (!pdlp_warm_start_data_) {
+    pdlp_warm_start_data_ = std::make_shared<pdlp_warm_start_data_t<i_t, f_t>>();
+  }
+  return *pdlp_warm_start_data_;
 }
 
+// These two live here rather than in solver_settings_accessors.cpp: they may have to
+// allocate the device-backed warm-start object, so they need CUDA.
 template <typename i_t, typename f_t>
-cpu_pdlp_warm_start_data_t<i_t, f_t>&
-pdlp_solver_settings_t<i_t, f_t>::get_cpu_pdlp_warm_start_data() noexcept
+const pdlp_warm_start_data_t<i_t, f_t>& pdlp_solver_settings_t<i_t, f_t>::get_pdlp_warm_start_data()
+  const noexcept
 {
-  return cpu_pdlp_warm_start_data_;
+  return ensure_pdlp_warm_start_data();
 }
 
 template <typename i_t, typename f_t>
-const pdlp_warm_start_data_view_t<i_t, f_t>&
-pdlp_solver_settings_t<i_t, f_t>::get_pdlp_warm_start_data_view() const noexcept
+pdlp_warm_start_data_t<i_t, f_t>& pdlp_solver_settings_t<i_t, f_t>::get_pdlp_warm_start_data()
 {
-  return pdlp_warm_start_data_view_;
+  return ensure_pdlp_warm_start_data();
 }
 
 #if MIP_INSTANTIATE_FLOAT || PDLP_INSTANTIATE_FLOAT

cpp/src/pdlp/solver_settings_accessors.cpp

@@ -0,0 +1,68 @@
+/* clang-format off */
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+/* clang-format on */
+
+// Warm-start accessors of pdlp_solver_settings_t, split out of solver_settings.cu.
+//
+// These are trivial `return member_;` getters -- they hand back a reference and emit no
+// device code, even where the referent is a GPU type. The gRPC client needs them, so they
+// build into the CUDA-free cuopt_client library while the rest of the class (which does
+// real thrust/rmm work) stays in solver_settings.cu.
+//
+// Only these members are instantiated below, deliberately NOT `template class`: the class
+// holds a pdlp_warm_start_data_t, so instantiating all of it here would pull in device
+// ctor/dtor code that belongs in the CUDA TU.
+
+#include <cuopt/export.hpp>
+#include <cuopt/mathematical_optimization/pdlp/solver_settings.hpp>
+
+// Required: the explicit instantiations below are guarded on MIP_INSTANTIATE_* /
+// PDLP_INSTANTIATE_*. Without this header those macros are undefined, the guards
+// evaluate false, and this TU silently compiles to zero symbols.
+#include <mip_heuristics/mip_constants.hpp>
+
+namespace cuopt::mathematical_optimization {
+
+template <typename i_t, typename f_t>
+const cpu_pdlp_warm_start_data_t<i_t, f_t>&
+pdlp_solver_settings_t<i_t, f_t>::get_cpu_pdlp_warm_start_data() const noexcept
+{
+  return cpu_pdlp_warm_start_data_;
+}
+
+template <typename i_t, typename f_t>
+cpu_pdlp_warm_start_data_t<i_t, f_t>&
+pdlp_solver_settings_t<i_t, f_t>::get_cpu_pdlp_warm_start_data() noexcept
+{
+  return cpu_pdlp_warm_start_data_;
+}
+
+template <typename i_t, typename f_t>
+const pdlp_warm_start_data_view_t<i_t, f_t>&
+pdlp_solver_settings_t<i_t, f_t>::get_pdlp_warm_start_data_view() const noexcept
+{
+  return pdlp_warm_start_data_view_;
+}
+
+#if MIP_INSTANTIATE_FLOAT || PDLP_INSTANTIATE_FLOAT
+template CUOPT_EXPORT const cpu_pdlp_warm_start_data_t<int, float>&
+pdlp_solver_settings_t<int, float>::get_cpu_pdlp_warm_start_data() const noexcept;
+template CUOPT_EXPORT cpu_pdlp_warm_start_data_t<int, float>&
+pdlp_solver_settings_t<int, float>::get_cpu_pdlp_warm_start_data() noexcept;
+template CUOPT_EXPORT const pdlp_warm_start_data_view_t<int, float>&
+pdlp_solver_settings_t<int, float>::get_pdlp_warm_start_data_view() const noexcept;
+#endif
+
+#if MIP_INSTANTIATE_DOUBLE
+template CUOPT_EXPORT const cpu_pdlp_warm_start_data_t<int, double>&
+pdlp_solver_settings_t<int, double>::get_cpu_pdlp_warm_start_data() const noexcept;
+template CUOPT_EXPORT cpu_pdlp_warm_start_data_t<int, double>&
+pdlp_solver_settings_t<int, double>::get_cpu_pdlp_warm_start_data() noexcept;
+template CUOPT_EXPORT const pdlp_warm_start_data_view_t<int, double>&
+pdlp_solver_settings_t<int, double>::get_pdlp_warm_start_data_view() const noexcept;
+#endif
+
+}  // namespace cuopt::mathematical_optimization