diff --git a/src/solver.cu b/src/solver.cu
index 2bdcb59..110b2cd 100644
--- a/src/solver.cu
+++ b/src/solver.cu
@@ -401,6 +401,7 @@ static pdhg_solver_state_t *initialize_solver_state(const lp_problem_t *working_
         (state->num_variables + state->num_constraints + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
     state->num_blocks_nnz = (state->constraint_matrix->num_nonzeros + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
 
+    CUDA_CHECK(cudaFree(0));
     CUSPARSE_CHECK(cusparseCreate(&state->sparse_handle));
     CUBLAS_CHECK(cublasCreate(&state->blas_handle));
     CUBLAS_CHECK(cublasSetPointerMode(state->blas_handle, CUBLAS_POINTER_MODE_HOST));
diff --git a/src/utils.cu b/src/utils.cu
index ef157db..bcf7e81 100644
--- a/src/utils.cu
+++ b/src/utils.cu
@@ -386,8 +386,8 @@ bool should_do_adaptive_restart(pdhg_solver_state_t *solver_state,
     return reason != NULL;
 }
 
-// The context is created once per process; a later CUDA call on any thread
-// waits for it inside the driver.
+// The solver thread must initialize the runtime before creating cuSPARSE
+// handles, which can otherwise race this background initialization.
 void cupdlpx_gpu_init_async(void)
 {
     static std::once_flag once;