Pdlp fix batch cuda graph (#68)

This PR aims at fixing the invalid operation while there is a graph
capture we sometimes see when using batch solve.

Solution is to use a regular instead of a non-blocking stream to make
sure that if any operation (like a cudaFree from Thrust) is being
launched on the default stream, it will wait for all other operations on
other stream to finish first, preventing any cudaMalloc/Free while
another stream might be doing a CUDA Graph capture.

Author: Kh4ster

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

@@ -103,7 +103,8 @@ struct solver_ret_t {
 // Wrapper for solve to expose the API to cython.
 
 std::unique_ptr<solver_ret_t> call_solve(cuopt::mps_parser::data_model_view_t<int, double>*,
-                                         linear_programming::solver_settings_t<int, double>*);
+                                         linear_programming::solver_settings_t<int, double>*,
+                                         unsigned int flags = cudaStreamNonBlocking);
 
 std::pair<std::vector<std::unique_ptr<solver_ret_t>>, double> call_batch_solve(
   std::vector<cuopt::mps_parser::data_model_view_t<int, double>*>,

cpp/src/linear_programming/utilities/cython_solve.cu

@@ -208,12 +208,13 @@ mip_ret_t call_solve_mip(
 
 std::unique_ptr<solver_ret_t> call_solve(
   cuopt::mps_parser::data_model_view_t<int, double>* data_model,
-  cuopt::linear_programming::solver_settings_t<int, double>* solver_settings)
+  cuopt::linear_programming::solver_settings_t<int, double>* solver_settings,
+  unsigned int flags)
 {
   raft::common::nvtx::range fun_scope("Call Solve");
 
   cudaStream_t stream;
-  RAFT_CUDA_TRY(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
+  RAFT_CUDA_TRY(cudaStreamCreateWithFlags(&stream, flags));
   const raft::handle_t handle_{stream};
 
   auto op_problem = data_model_to_optimization_problem(data_model, solver_settings, &handle_);
@@ -283,9 +284,11 @@ std::pair<std::vector<std::unique_ptr<solver_ret_t>>, double> call_batch_solve(
     solver_settings->set_parameter(CUOPT_METHOD, CUOPT_METHOD_PDLP);
   }
 
+  // Use a default stream instead of a non-blocking to avoid invalid operations while some CUDA
+  // Graph might be capturing in another stream
 #pragma omp parallel for num_threads(max_thread)
   for (std::size_t i = 0; i < size; ++i)
-    list[i] = std::move(call_solve(data_models[i], solver_settings));
+    list[i] = std::move(call_solve(data_models[i], solver_settings, cudaStreamDefault));
 
   auto end      = std::chrono::high_resolution_clock::now();
   auto duration = std::chrono::duration_cast<std::chrono::milliseconds>(end - start_solver);