Only run cuda device on rank 0

src/core/scatter_devices/scatter_device_factory.cpp

@@ -156,10 +156,9 @@ ScatterDeviceFactory::create(boost::mpi::communicator &scatter_comm, Sample &sam
     // all_comm for inter-partition communication, parition_comm for
     // intra-partition communication
     if (Params::Inst()->scattering.type == "self") {
-        if (Params::Inst()->device == sass::platform::Device::CUDA) {
-            if (scatter_comm.rank() == 0) {
-                Info::Inst()->write("Initializing CUDA Scatter Device, Vectors (self)");
-            }
+        // Only rank 0 should claim the CUDA device
+        if (Params::Inst()->device == sass::platform::Device::CUDA && scatter_comm.rank() == 0) {
+            Info::Inst()->write("Initializing CUDA Scatter Device, Vectors (self)");
             p_ScatterDevice = new SelfVectorsCUDAScatterDevice(
                 all_comm, partition_comm, sample, thispartition_QIV, NAF, fileservice_endpoint,
                 monitorservice_endpoint);