diff --git a/docs/technologies/eigensolvers/cusolvermp.md b/docs/technologies/eigensolvers/cusolvermp.md
index eb46dc0167..ac59195bdd 100644
--- a/docs/technologies/eigensolvers/cusolvermp.md
+++ b/docs/technologies/eigensolvers/cusolvermp.md
@@ -34,8 +34,9 @@ diagonalization paths that avoid a CP2K-side Cholesky reduction where the select
 supports them. With [cuSOLVERMp], this currently covers real symmetric and complex Hermitian
 generalized eigenproblems through `cusolverMpSygvd`.
 
-When [cuSOLVERMp] uses the NCCL backend, CP2K requires at most one local MPI rank per visible GPU.
-Use one MPI rank per GPU or reduce the number of local MPI ranks when only one GPU is visible.
+When [cuSOLVERMp] uses the NCCL backend, CP2K relies on the launcher or scheduler to bind MPI ranks
+to GPUs. For example, assigning one GPU per rank with `CUDA_VISIBLE_DEVICES` is supported. CP2K only
+checks that the CUDA device selected for each MPI rank is visible to that rank.
 
 [cal]: https://developer.download.nvidia.com/compute/cublasmp/redist/libcal/
 [cusolvermp]: https://docs.nvidia.com/cuda/cusolvermp/
diff --git a/src/fm/cp_fm_cusolver.c b/src/fm/cp_fm_cusolver.c
index b4e6ce7497..159b76a200 100644
--- a/src/fm/cp_fm_cusolver.c
+++ b/src/fm/cp_fm_cusolver.c
@@ -54,57 +54,31 @@
   } while (0)
 
 /*******************************************************************************
- * \brief Check that local MPI ranks can be mapped one-to-one to GPUs.
+ * \brief Check that the selected CUDA device is visible to this MPI rank.
  ******************************************************************************/
-static void check_nccl_rank_device_mapping(MPI_Comm comm,
-                                           const int local_device) {
+static void check_nccl_device_visibility(MPI_Comm comm,
+                                         const int local_device) {
   int device_count = 0;
   CUDA_CHECK(cudaGetDeviceCount(&device_count));
 
-  MPI_Comm node_comm;
-  MPI_Comm_split_type(comm, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL, &node_comm);
-
-  int rank, local_rank, local_nranks;
+  int rank;
   MPI_Comm_rank(comm, &rank);
-  MPI_Comm_rank(node_comm, &local_rank);
-  MPI_Comm_size(node_comm, &local_nranks);
 
   if (device_count <= 0) {
-    if (local_rank == 0) {
-      fprintf(stderr,
-              "ERROR: cuSOLVERMp with NCCL requires at least one visible GPU "
-              "on every node. No CUDA device is visible on the node containing "
-              "MPI rank %d.\n",
-              rank);
-    }
-    MPI_Comm_free(&node_comm);
+    fprintf(stderr,
+            "ERROR: cuSOLVERMp with NCCL requires at least one visible GPU. "
+            "No CUDA device is visible to MPI rank %d.\n",
+            rank);
     MPI_Abort(comm, EXIT_FAILURE);
   }
 
   if (local_device < 0 || local_device >= device_count) {
     fprintf(stderr,
             "ERROR: cuSOLVERMp with NCCL selected invalid CUDA device %d on "
-            "MPI rank %d; %d CUDA device(s) are visible on this node.\n",
+            "MPI rank %d; %d CUDA device(s) are visible to this rank.\n",
             local_device, rank, device_count);
-    MPI_Comm_free(&node_comm);
     MPI_Abort(comm, EXIT_FAILURE);
   }
-
-  if (local_nranks > device_count) {
-    if (local_rank == 0) {
-      fprintf(stderr,
-              "ERROR: cuSOLVERMp with NCCL in CP2K requires at most one local "
-              "MPI rank per visible GPU. This node has %d local MPI ranks and "
-              "%d visible GPU(s). Use fewer MPI ranks per node, expose more "
-              "GPUs with CUDA_VISIBLE_DEVICES, or choose another "
-              "PREFERRED_DIAG_LIBRARY.\n",
-              local_nranks, device_count);
-    }
-    MPI_Comm_free(&node_comm);
-    MPI_Abort(comm, EXIT_FAILURE);
-  }
-
-  MPI_Comm_free(&node_comm);
 }
 
 #else
@@ -279,7 +253,7 @@ void cp_fm_diag_cusolver(const int fortran_comm, const int matrix_desc[9],
   MPI_Comm_size(comm, &nranks);
 
 #if defined(__CUSOLVERMP_NCCL)
-  check_nccl_rank_device_mapping(comm, local_device);
+  check_nccl_device_visibility(comm, local_device);
 #endif
   offload_activate_chosen_device();
 
@@ -444,7 +418,7 @@ void cp_fm_diag_cusolver_sygvd(const int fortran_comm,
   MPI_Comm_size(comm, &nranks);
 
 #if defined(__CUSOLVERMP_NCCL)
-  check_nccl_rank_device_mapping(comm, local_device);
+  check_nccl_device_visibility(comm, local_device);
 #endif
   offload_activate_chosen_device();
 
@@ -646,7 +620,7 @@ void cp_cfm_diag_cusolver_hegvd(
   MPI_Comm_size(comm, &nranks);
 
 #if defined(__CUSOLVERMP_NCCL)
-  check_nccl_rank_device_mapping(comm, local_device);
+  check_nccl_device_visibility(comm, local_device);
 #endif
   offload_activate_chosen_device();