diff --git a/packages/stokhos/src/sacado/kokkos/pce/amesos2/Amesos2_Solver_UQ_PCE.hpp b/packages/stokhos/src/sacado/kokkos/pce/amesos2/Amesos2_Solver_UQ_PCE.hpp
index db3718af3759..1927b2aed9b8 100644
--- a/packages/stokhos/src/sacado/kokkos/pce/amesos2/Amesos2_Solver_UQ_PCE.hpp
+++ b/packages/stokhos/src/sacado/kokkos/pce/amesos2/Amesos2_Solver_UQ_PCE.hpp
@@ -49,12 +49,12 @@
 
 namespace Amesos2 {
 
-  template <class S, class LO, class GO, class D>
+  template <class S, class LO, class GO, class NO>
   typename Sacado::UQ::PCE<S>::cijk_type
   get_pce_cijk(
-    const Teuchos::RCP<const Tpetra::CrsMatrix<Sacado::UQ::PCE<S>, LO, GO, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<D> > >& A = Teuchos::null,
-    const Teuchos::RCP<Tpetra::MultiVector<Sacado::UQ::PCE<S>, LO, GO, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<D> > >& X = Teuchos::null,
-    const Teuchos::RCP<const Tpetra::MultiVector<Sacado::UQ::PCE<S>, LO, GO, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<D> > >& B = Teuchos::null)
+    const Teuchos::RCP<const Tpetra::CrsMatrix<Sacado::UQ::PCE<S>, LO, GO, NO > >& A = Teuchos::null,
+    const Teuchos::RCP<Tpetra::MultiVector<Sacado::UQ::PCE<S>, LO, GO, NO > >& X = Teuchos::null,
+    const Teuchos::RCP<const Tpetra::MultiVector<Sacado::UQ::PCE<S>, LO, GO, NO > >& B = Teuchos::null)
   {
     if (A != Teuchos::null) {
       return Kokkos::cijk(A->getLocalValuesDevice(Tpetra::Access::ReadOnly));
@@ -75,22 +75,21 @@ namespace Amesos2 {
   /// these matrices and vectors into ones with a standard (e.g., double)
   /// scalar type.
   template <class Storage, class LocalOrdinal, class GlobalOrdinal,
-            class Device, template<class,class> class ConcreteSolver>
+            class Node, template<class,class> class ConcreteSolver>
   class PCESolverAdapter :
     public Solver< Tpetra::CrsMatrix<Sacado::UQ::PCE<Storage>,
                                      LocalOrdinal,
                                      GlobalOrdinal,
-                                     Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> >,
+                                     Node >,
                    Tpetra::MultiVector<Sacado::UQ::PCE<Storage>,
                                        LocalOrdinal,
                                        GlobalOrdinal,
-                                       Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> >
+                                       Node >
                    >
   {
   public:
 
     typedef Sacado::UQ::PCE<Storage> Scalar;
-    typedef Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> Node;
     typedef Tpetra::CrsMatrix<Scalar,LocalOrdinal,GlobalOrdinal,Node> Matrix;
     typedef Tpetra::MultiVector<Scalar,LocalOrdinal,GlobalOrdinal,Node> Vector;
 
@@ -509,13 +508,12 @@ namespace Amesos2 {
   // Sacado::UQ::PCE where we create PCESolverAdapter wrapping
   // each solver
   template < template <class,class> class ConcreteSolver,
-             class ST, class LO, class GO, class D >
+             class ST, class LO, class GO, class NO >
   struct create_solver_with_supported_type<
     ConcreteSolver,
-    Tpetra::CrsMatrix<Sacado::UQ::PCE<ST>,LO,GO,Tpetra::KokkosCompat::KokkosDeviceWrapperNode<D> >,
-    Tpetra::MultiVector<Sacado::UQ::PCE<ST>,LO,GO,Tpetra::KokkosCompat::KokkosDeviceWrapperNode<D> > > {
+    Tpetra::CrsMatrix<Sacado::UQ::PCE<ST>,LO,GO,NO >,
+    Tpetra::MultiVector<Sacado::UQ::PCE<ST>,LO,GO,NO > > {
     typedef Sacado::UQ::PCE<ST> SC;
-    typedef Tpetra::KokkosCompat::KokkosDeviceWrapperNode<D> NO;
     typedef Tpetra::CrsMatrix<SC,LO,GO,NO> Matrix;
     typedef Tpetra::MultiVector<SC,LO,GO,NO> Vector;
     static Teuchos::RCP<Solver<Matrix,Vector> >
@@ -532,7 +530,7 @@ namespace Amesos2 {
       (void)same_scalar_assertion; // This stops the compiler from warning about unused declared variables
 
       // If our assertion did not fail, then create and return a new solver
-      return Teuchos::rcp( new PCESolverAdapter<ST,LO,GO,D,ConcreteSolver>(A, X, B) );
+      return Teuchos::rcp( new PCESolverAdapter<ST,LO,GO,NO,ConcreteSolver>(A, X, B) );
     }
   };
 
diff --git a/packages/stokhos/src/sacado/kokkos/pce/tpetra/Stokhos_Tpetra_UQ_PCE.hpp b/packages/stokhos/src/sacado/kokkos/pce/tpetra/Stokhos_Tpetra_UQ_PCE.hpp
index 3a1bd60faa4b..c3e1a94d0235 100644
--- a/packages/stokhos/src/sacado/kokkos/pce/tpetra/Stokhos_Tpetra_UQ_PCE.hpp
+++ b/packages/stokhos/src/sacado/kokkos/pce/tpetra/Stokhos_Tpetra_UQ_PCE.hpp
@@ -127,9 +127,9 @@ struct DeviceForNode2 {
 #endif // defined(KOKKOS_ENABLE_SERIAL)
 };
 
-template <typename Device>
-struct DeviceForNode2< Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > {
-  typedef Device type;
+template <typename ExecSpace, typename MemSpace>
+struct DeviceForNode2< Tpetra::KokkosCompat::KokkosDeviceWrapperNode<ExecSpace, MemSpace> > {
+  typedef typename Tpetra::KokkosCompat::KokkosDeviceWrapperNode<ExecSpace, MemSpace>::device_type type;
 };
 
 }
diff --git a/packages/stokhos/src/sacado/kokkos/pce/tpetra/Stokhos_Tpetra_Utilities_UQ_PCE.hpp b/packages/stokhos/src/sacado/kokkos/pce/tpetra/Stokhos_Tpetra_Utilities_UQ_PCE.hpp
index e82d6ac22329..9c47cf03db7d 100644
--- a/packages/stokhos/src/sacado/kokkos/pce/tpetra/Stokhos_Tpetra_Utilities_UQ_PCE.hpp
+++ b/packages/stokhos/src/sacado/kokkos/pce/tpetra/Stokhos_Tpetra_Utilities_UQ_PCE.hpp
@@ -53,17 +53,15 @@ namespace Stokhos {
 
 
   // Build a CRS graph from a sparse Cijk tensor
-  template <typename LocalOrdinal, typename GlobalOrdinal, typename Device,
+  template <typename LocalOrdinal, typename GlobalOrdinal, typename Node,
             typename CijkType>
-  Teuchos::RCP< Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,
-                                 Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >
+  Teuchos::RCP< Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Node > >
   create_cijk_crs_graph(const CijkType& cijk_dev,
                         const Teuchos::RCP<const Teuchos::Comm<int> >& comm,
                         const size_t matrix_pce_size) {
     using Teuchos::RCP;
     using Teuchos::arrayView;
 
-    typedef Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> Node;
     typedef Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node> Map;
     typedef Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Node> Graph;
 
@@ -117,16 +115,15 @@ namespace Stokhos {
   // UQ::PCE scalar type
   // If flat_domain_map and/or flat_range_map are null, they will be computed,
   // otherwise they will be used as-is.
-  template <typename LocalOrdinal, typename GlobalOrdinal, typename Device,
+  template <typename LocalOrdinal, typename GlobalOrdinal, typename Node,
             typename CijkType>
-  Teuchos::RCP< Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,
-                                 Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >
+  Teuchos::RCP< Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Node > >
   create_flat_pce_graph(
-    const Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> >& graph,
+    const Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Node >& graph,
     const CijkType& cijk,
-    Teuchos::RCP<const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >& flat_domain_map,
-    Teuchos::RCP<const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >& flat_range_map,
-    Teuchos::RCP<const Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >& cijk_graph,
+    Teuchos::RCP<const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node > >& flat_domain_map,
+    Teuchos::RCP<const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node > >& flat_range_map,
+    Teuchos::RCP<const Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Node > >& cijk_graph,
     const size_t matrix_pce_size) {
     using Teuchos::ArrayView;
     using Teuchos::ArrayRCP;
@@ -134,7 +131,6 @@ namespace Stokhos {
     using Teuchos::RCP;
     using Teuchos::rcp;
 
-    typedef Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> Node;
     typedef Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node> Map;
     typedef Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Node> Graph;
 
@@ -163,7 +159,7 @@ namespace Stokhos {
 
     // Build Cijk graph if necessary
     if (cijk_graph == Teuchos::null)
-      cijk_graph = create_cijk_crs_graph<LocalOrdinal,GlobalOrdinal,Device>(
+      cijk_graph = create_cijk_crs_graph<LocalOrdinal,GlobalOrdinal,Node>(
         cijk,
         flat_domain_map->getComm(),
         matrix_pce_size);
@@ -231,21 +227,17 @@ namespace Stokhos {
   // Create a flattened vector by unrolling the UQ::PCE scalar type.  The
   // returned vector is a view of the original
   template <typename Storage, typename LocalOrdinal, typename GlobalOrdinal,
-            typename Device>
+            typename Node>
   Teuchos::RCP< const Tpetra::MultiVector<typename Storage::value_type,
-                                          LocalOrdinal,GlobalOrdinal,
-                                          Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >
+                                          LocalOrdinal,GlobalOrdinal,Node > >
   create_flat_vector_view(
     const Tpetra::MultiVector<Sacado::UQ::PCE<Storage>,
-                              LocalOrdinal,GlobalOrdinal,
-                              Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> >& vec,
-    const Teuchos::RCP< const Tpetra::Map<LocalOrdinal,GlobalOrdinal,
-                                          Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >& flat_map) {
+                              LocalOrdinal,GlobalOrdinal,Node >& vec,
+    const Teuchos::RCP< const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node > >& flat_map) {
     using Teuchos::RCP;
     using Teuchos::rcp;
 
     typedef typename Storage::value_type BaseScalar;
-    typedef Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> Node;
     typedef Tpetra::MultiVector<BaseScalar,LocalOrdinal,GlobalOrdinal,Node> FlatVector;
     typedef typename FlatVector::dual_view_type::t_dev flat_view_type;
 
@@ -253,7 +245,7 @@ namespace Stokhos {
     // non-const method, yet getLocalViewDevice(ReadOnly) returns a const-view
     // (i.e., with a constant scalar type), and there is no way to make a
     // MultiVector out of it!
-    typedef Tpetra::MultiVector<Sacado::UQ::PCE<Storage>, LocalOrdinal,GlobalOrdinal, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > mv_type;
+    typedef Tpetra::MultiVector<Sacado::UQ::PCE<Storage>, LocalOrdinal,GlobalOrdinal, Node > mv_type;
     mv_type& vec_nc = const_cast<mv_type&>(vec);
 
     // Create flattenend view using special reshaping view assignment operator
@@ -268,21 +260,17 @@ namespace Stokhos {
   // Create a flattened vector by unrolling the UQ::PCE scalar type.  The
   // returned vector is a view of the original
   template <typename Storage, typename LocalOrdinal, typename GlobalOrdinal,
-            typename Device>
+            typename Node>
   Teuchos::RCP< Tpetra::MultiVector<typename Storage::value_type,
-                                    LocalOrdinal,GlobalOrdinal,
-                                    Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >
+                                    LocalOrdinal,GlobalOrdinal,Node > >
   create_flat_vector_view(
     Tpetra::MultiVector<Sacado::UQ::PCE<Storage>,
-                        LocalOrdinal,GlobalOrdinal,
-                        Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> >& vec,
-    const Teuchos::RCP< const Tpetra::Map<LocalOrdinal,GlobalOrdinal,
-                                          Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >& flat_map) {
+                        LocalOrdinal,GlobalOrdinal,Node >& vec,
+    const Teuchos::RCP< const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node > >& flat_map) {
     using Teuchos::RCP;
     using Teuchos::rcp;
 
     typedef typename Storage::value_type BaseScalar;
-    typedef Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> Node;
     typedef Tpetra::MultiVector<BaseScalar,LocalOrdinal,GlobalOrdinal,Node> FlatVector;
     typedef typename FlatVector::dual_view_type::t_dev flat_view_type;
 
@@ -299,17 +287,14 @@ namespace Stokhos {
   // returned vector is a view of the original.  This version creates the
   // map if necessary
   template <typename Storage, typename LocalOrdinal, typename GlobalOrdinal,
-            typename Device>
+            typename Node>
   Teuchos::RCP< const Tpetra::MultiVector<typename Storage::value_type,
-                                          LocalOrdinal,GlobalOrdinal,
-                                          Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >
+                                          LocalOrdinal,GlobalOrdinal,Node > >
   create_flat_vector_view(
     const Tpetra::MultiVector<Sacado::UQ::PCE<Storage>,
-                              LocalOrdinal,GlobalOrdinal,
-                              Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> >& vec,
-    Teuchos::RCP< const Tpetra::Map<LocalOrdinal,GlobalOrdinal,
-                                    Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >& flat_map) {
-    typedef Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> Node;
+                              LocalOrdinal,GlobalOrdinal,Node >& vec,
+    Teuchos::RCP< const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node > >& flat_map) {
+    typedef typename Node::device_type Device;
     if (flat_map == Teuchos::null) {
       const LocalOrdinal pce_size =
         Kokkos::dimension_scalar(vec.template getLocalView<Device>(Tpetra::Access::ReadOnly));
@@ -323,17 +308,14 @@ namespace Stokhos {
   // returned vector is a view of the original.  This version creates the
   // map if necessary
   template <typename Storage, typename LocalOrdinal, typename GlobalOrdinal,
-            typename Device>
+            typename Node>
   Teuchos::RCP< Tpetra::MultiVector<typename Storage::value_type,
-                                    LocalOrdinal,GlobalOrdinal,
-                                    Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >
+                                    LocalOrdinal,GlobalOrdinal,Node > >
   create_flat_vector_view(
     Tpetra::MultiVector<Sacado::UQ::PCE<Storage>,
-                        LocalOrdinal,GlobalOrdinal,
-                        Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> >& vec,
-    Teuchos::RCP< const Tpetra::Map<LocalOrdinal,GlobalOrdinal,
-                                    Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >& flat_map) {
-    typedef Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> Node;
+                        LocalOrdinal,GlobalOrdinal,Node >& vec,
+    Teuchos::RCP< const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node > >& flat_map) {
+    typedef typename Node::device_type Device;
     if (flat_map == Teuchos::null) {
       const LocalOrdinal pce_size =
         Kokkos::dimension_scalar(vec.template getLocalView<Device>(Tpetra::Access::ReadOnly));
@@ -346,17 +328,13 @@ namespace Stokhos {
   // Create a flattened vector by unrolling the UQ::PCE scalar type.  The
   // returned vector is a view of the original
   template <typename Storage, typename LocalOrdinal, typename GlobalOrdinal,
-            typename Device>
+            typename Node>
   Teuchos::RCP< const Tpetra::Vector<typename Storage::value_type,
-                                     LocalOrdinal,GlobalOrdinal,
-                                     Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >
+                                     LocalOrdinal,GlobalOrdinal,Node > >
   create_flat_vector_view(
     const Tpetra::Vector<Sacado::UQ::PCE<Storage>,
-                         LocalOrdinal,GlobalOrdinal,
-                         Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> >& vec_const,
-    const Teuchos::RCP< const Tpetra::Map<LocalOrdinal,GlobalOrdinal,
-                                          Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >& flat_map) {
-    typedef Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> Node;
+                         LocalOrdinal,GlobalOrdinal,Node >& vec_const,
+    const Teuchos::RCP< const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node > >& flat_map) {
     const Tpetra::MultiVector<Sacado::UQ::PCE<Storage>,LocalOrdinal,GlobalOrdinal,Node>& mv = vec_const;
     Teuchos::RCP< Tpetra::MultiVector<typename Storage::value_type,LocalOrdinal,GlobalOrdinal,Node> > flat_mv = create_flat_vector_view(mv, flat_map);
     return flat_mv->getVector(0);
@@ -366,17 +344,14 @@ namespace Stokhos {
   // returned vector is a view of the original.  This version creates the
   // map if necessary
   template <typename Storage, typename LocalOrdinal, typename GlobalOrdinal,
-            typename Device>
+            typename Node>
   Teuchos::RCP< const Tpetra::Vector<typename Storage::value_type,
-                                     LocalOrdinal,GlobalOrdinal,
-                                     Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >
+                                     LocalOrdinal,GlobalOrdinal,Node > >
   create_flat_vector_view(
     const Tpetra::Vector<Sacado::UQ::PCE<Storage>,
-                         LocalOrdinal,GlobalOrdinal,
-                         Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> >& vec,
-    Teuchos::RCP< const Tpetra::Map<LocalOrdinal,GlobalOrdinal,
-                                    Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >& flat_map) {
-    typedef Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> Node;
+                         LocalOrdinal,GlobalOrdinal,Node >& vec,
+    Teuchos::RCP< const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node > >& flat_map) {
+    typedef typename Node::device_type Device;
     if (flat_map == Teuchos::null) {
       const LocalOrdinal pce_size =
         Kokkos::dimension_scalar(vec.template getLocalView<Device>(Tpetra::Access::ReadOnly));
@@ -389,17 +364,13 @@ namespace Stokhos {
   // Create a flattened vector by unrolling the UQ::PCE scalar type.  The
   // returned vector is a view of the original
   template <typename Storage, typename LocalOrdinal, typename GlobalOrdinal,
-            typename Device>
+            typename Node>
   Teuchos::RCP< Tpetra::Vector<typename Storage::value_type,
-                               LocalOrdinal,GlobalOrdinal,
-                               Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >
+                               LocalOrdinal,GlobalOrdinal,Node > >
   create_flat_vector_view(
     Tpetra::Vector<Sacado::UQ::PCE<Storage>,
-                   LocalOrdinal,GlobalOrdinal,
-                   Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> >& vec,
-    const Teuchos::RCP< const Tpetra::Map<LocalOrdinal,GlobalOrdinal,
-                                          Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >& flat_map) {
-    typedef Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> Node;
+                   LocalOrdinal,GlobalOrdinal,Node >& vec,
+    const Teuchos::RCP< const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node > >& flat_map) {
     Tpetra::MultiVector<Sacado::UQ::PCE<Storage>,LocalOrdinal,GlobalOrdinal,Node>& mv = vec;
     Teuchos::RCP< Tpetra::MultiVector<typename Storage::value_type,LocalOrdinal,GlobalOrdinal,Node> > flat_mv = create_flat_vector_view(mv, flat_map);
     return flat_mv->getVectorNonConst(0);
@@ -409,17 +380,14 @@ namespace Stokhos {
   // returned vector is a view of the original.  This version creates the
   // map if necessary
   template <typename Storage, typename LocalOrdinal, typename GlobalOrdinal,
-            typename Device>
+            typename Node>
   Teuchos::RCP< Tpetra::Vector<typename Storage::value_type,
-                               LocalOrdinal,GlobalOrdinal,
-                               Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >
+                               LocalOrdinal,GlobalOrdinal,Node > >
   create_flat_vector_view(
     Tpetra::Vector<Sacado::UQ::PCE<Storage>,
-                   LocalOrdinal,GlobalOrdinal,
-                   Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> >& vec,
-    Teuchos::RCP< const Tpetra::Map<LocalOrdinal,GlobalOrdinal,
-                                    Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >& flat_map) {
-    typedef Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> Node;
+                   LocalOrdinal,GlobalOrdinal,Node >& vec,
+    Teuchos::RCP< const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node > >& flat_map) {
+    typedef typename Node::device_type Device;
     if (flat_map == Teuchos::null) {
       const LocalOrdinal pce_size =
         Kokkos::dimension_scalar(vec.template getLocalView<Device>(Tpetra::Access::ReadOnly));
@@ -432,22 +400,20 @@ namespace Stokhos {
   // Create a flattened matrix by unrolling the UQ::PCE scalar type.  The
   // returned matrix is NOT a view of the original (and can't be)
   template <typename Storage, typename LocalOrdinal, typename GlobalOrdinal,
-            typename Device, typename CijkType>
+            typename Node, typename CijkType>
   Teuchos::RCP< Tpetra::CrsMatrix<typename Storage::value_type,
-                                  LocalOrdinal,GlobalOrdinal,
-                                  Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >
+                                  LocalOrdinal,GlobalOrdinal,Node > >
   create_flat_matrix(
     const Tpetra::CrsMatrix<Sacado::UQ::PCE<Storage>,
-                            LocalOrdinal,GlobalOrdinal,Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> >& mat,
-    const Teuchos::RCP<const Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >& flat_graph,
-    const Teuchos::RCP<const Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >& cijk_graph,
+                            LocalOrdinal,GlobalOrdinal,Node >& mat,
+    const Teuchos::RCP<const Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Node > >& flat_graph,
+    const Teuchos::RCP<const Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Node > >& cijk_graph,
     const CijkType& cijk_dev) {
     using Teuchos::ArrayView;
     using Teuchos::Array;
     using Teuchos::RCP;
     using Teuchos::rcp;
 
-    typedef Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> Node;
     typedef Sacado::UQ::PCE<Storage> Scalar;
     typedef typename Storage::value_type BaseScalar;
     typedef Tpetra::CrsMatrix<Scalar,LocalOrdinal,GlobalOrdinal,Node> Matrix;
diff --git a/packages/stokhos/src/sacado/kokkos/vector/amesos2/Amesos2_Solver_MP_Vector.hpp b/packages/stokhos/src/sacado/kokkos/vector/amesos2/Amesos2_Solver_MP_Vector.hpp
index db434a952bea..87813021f05b 100644
--- a/packages/stokhos/src/sacado/kokkos/vector/amesos2/Amesos2_Solver_MP_Vector.hpp
+++ b/packages/stokhos/src/sacado/kokkos/vector/amesos2/Amesos2_Solver_MP_Vector.hpp
@@ -49,21 +49,11 @@
 
 namespace Amesos2 {
 
-  template <class S, class LO, class GO, class N>
+  template <class S, class LO, class GO, class NO>
   LO get_mp_vector_size(
-    const Teuchos::RCP<const Tpetra::CrsMatrix<Sacado::MP::Vector<S>, LO, GO, N> >& A = Teuchos::null,
-    const Teuchos::RCP<Tpetra::MultiVector<Sacado::MP::Vector<S>, LO, GO, N> >& X = Teuchos::null,
-    const Teuchos::RCP<const Tpetra::MultiVector<Sacado::MP::Vector<S>, LO, GO, N> >& B = Teuchos::null)
-  {
-    // Without KokkosRefactor, can only do static
-    return S::static_size;
-  }
-
-  template <class S, class LO, class GO, class D>
-  LO get_mp_vector_size(
-    const Teuchos::RCP<const Tpetra::CrsMatrix<Sacado::MP::Vector<S>, LO, GO, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<D> > >& A = Teuchos::null,
-    const Teuchos::RCP<Tpetra::MultiVector<Sacado::MP::Vector<S>, LO, GO, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<D> > >& X = Teuchos::null,
-    const Teuchos::RCP<const Tpetra::MultiVector<Sacado::MP::Vector<S>, LO, GO, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<D> > >& B = Teuchos::null)
+    const Teuchos::RCP<const Tpetra::CrsMatrix<Sacado::MP::Vector<S>, LO, GO, NO > >& A = Teuchos::null,
+    const Teuchos::RCP<Tpetra::MultiVector<Sacado::MP::Vector<S>, LO, GO, NO > >& X = Teuchos::null,
+    const Teuchos::RCP<const Tpetra::MultiVector<Sacado::MP::Vector<S>, LO, GO, NO > >& B = Teuchos::null)
   {
     if (A != Teuchos::null) {
       return Kokkos::dimension_scalar(A->getLocalValuesDevice(Tpetra::Access::ReadOnly));
diff --git a/packages/stokhos/src/sacado/kokkos/vector/linalg/Kokkos_CrsMatrix_MP_Vector_Cuda.hpp b/packages/stokhos/src/sacado/kokkos/vector/linalg/Kokkos_CrsMatrix_MP_Vector_Cuda.hpp
index 4e29e05c68a3..a1df5529096c 100644
--- a/packages/stokhos/src/sacado/kokkos/vector/linalg/Kokkos_CrsMatrix_MP_Vector_Cuda.hpp
+++ b/packages/stokhos/src/sacado/kokkos/vector/linalg/Kokkos_CrsMatrix_MP_Vector_Cuda.hpp
@@ -78,7 +78,8 @@ FullOccupancyKernelLaunch(Kernel kernel) {
 //
 // This implementation uses the underlying 2-D view directly.
 // Currently only works for statically sized MP::Vector
-template <typename MatrixStorage,
+template <typename MatrixMemorySpace,
+          typename MatrixStorage,
           typename MatrixOrdinal,
           typename MatrixMemory,
           typename MatrixSize,
@@ -89,7 +90,7 @@ template <typename MatrixStorage,
           typename Update>
 class MPMultiply< KokkosSparse::CrsMatrix<const Sacado::MP::Vector<MatrixStorage>,
                                     MatrixOrdinal,
-                                    Kokkos::Device<Kokkos::Cuda, typename Kokkos::Cuda::memory_space>,
+                                    Kokkos::Device<Kokkos::Cuda, MatrixMemorySpace>,
                                     MatrixMemory,
                                     MatrixSize>,
                   Kokkos::View< const Sacado::MP::Vector<InputStorage>*,
@@ -106,7 +107,7 @@ class MPMultiply< KokkosSparse::CrsMatrix<const Sacado::MP::Vector<MatrixStorage
   typedef Sacado::MP::Vector<InputStorage> InputVectorValue;
   typedef Sacado::MP::Vector<OutputStorage> OutputVectorValue;
 
-  typedef Kokkos::Device<Kokkos::Cuda, typename Kokkos::Cuda::memory_space> Device;
+  typedef Kokkos::Device<Kokkos::Cuda, MatrixMemorySpace> Device;
   typedef typename Device::execution_space execution_space;
   typedef typename execution_space::size_type size_type;
 
@@ -325,7 +326,8 @@ class MPMultiply< KokkosSparse::CrsMatrix<const Sacado::MP::Vector<MatrixStorage
 //
 // This implementation uses the underlying 2-D view directly.
 // Currently only works for statically sized MP::Vector
-template <typename MatrixStorage,
+template <typename MatrixMemorySpace,
+          typename MatrixStorage,
           typename MatrixOrdinal,
           typename MatrixMemory,
           typename MatrixSize,
@@ -336,7 +338,7 @@ template <typename MatrixStorage,
           typename Update>
 class MPMultiply< KokkosSparse::CrsMatrix<const Sacado::MP::Vector<MatrixStorage>,
                                     MatrixOrdinal,
-                                    Kokkos::Device<Kokkos::Cuda, typename Kokkos::Cuda::memory_space>,
+                                    Kokkos::Device<Kokkos::Cuda, MatrixMemorySpace>,
                                     MatrixMemory,
                                     MatrixSize>,
                   Kokkos::View< const Sacado::MP::Vector<InputStorage>**,
@@ -353,7 +355,7 @@ class MPMultiply< KokkosSparse::CrsMatrix<const Sacado::MP::Vector<MatrixStorage
   typedef Sacado::MP::Vector<InputStorage> InputVectorValue;
   typedef Sacado::MP::Vector<OutputStorage> OutputVectorValue;
 
-  typedef Kokkos::Device<Kokkos::Cuda, typename Kokkos::Cuda::memory_space> Device;
+  typedef Kokkos::Device<Kokkos::Cuda, MatrixMemorySpace> Device;
   typedef typename Device::execution_space execution_space;
   typedef typename execution_space::size_type size_type;
 
diff --git a/packages/stokhos/src/sacado/kokkos/vector/tpetra/Stokhos_Tpetra_MP_Vector.hpp b/packages/stokhos/src/sacado/kokkos/vector/tpetra/Stokhos_Tpetra_MP_Vector.hpp
index 7fed0dd22b63..72a8cefb0218 100644
--- a/packages/stokhos/src/sacado/kokkos/vector/tpetra/Stokhos_Tpetra_MP_Vector.hpp
+++ b/packages/stokhos/src/sacado/kokkos/vector/tpetra/Stokhos_Tpetra_MP_Vector.hpp
@@ -101,9 +101,9 @@ struct DeviceForNode {
 /// \brief Partial specialization of DeviceForNode, for new Kokkos
 ///   "wrapper" Node types.
 /// \tparam Device (New) Kokkos execution space type.
-template <typename Device>
-struct DeviceForNode< Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > {
-  typedef Device type;
+template <typename ExecSpace, typename MemSpace>
+struct DeviceForNode< Tpetra::KokkosCompat::KokkosDeviceWrapperNode<ExecSpace, MemSpace> > {
+  typedef typename Tpetra::KokkosCompat::KokkosDeviceWrapperNode<ExecSpace, MemSpace>::device_type type;
 };
 #endif
 
diff --git a/packages/stokhos/src/sacado/kokkos/vector/tpetra/Stokhos_Tpetra_Utilities_MP_Vector.hpp b/packages/stokhos/src/sacado/kokkos/vector/tpetra/Stokhos_Tpetra_Utilities_MP_Vector.hpp
index 3370710d3969..360304b0a9fc 100644
--- a/packages/stokhos/src/sacado/kokkos/vector/tpetra/Stokhos_Tpetra_Utilities_MP_Vector.hpp
+++ b/packages/stokhos/src/sacado/kokkos/vector/tpetra/Stokhos_Tpetra_Utilities_MP_Vector.hpp
@@ -93,91 +93,16 @@ namespace Stokhos {
   // If flat_domain_map and/or flat_range_map are null, they will be computed,
   // otherwise they will be used as-is.
   template <typename LocalOrdinal, typename GlobalOrdinal, typename Node>
-  Teuchos::RCP< Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Node> >
+  Teuchos::RCP< Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Node > >
   create_flat_mp_graph(
-    const Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Node>& graph,
-    Teuchos::RCP<const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node> >& flat_domain_map,
-    Teuchos::RCP<const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node> >& flat_range_map,
+    const Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Node >& graph,
+    Teuchos::RCP<const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node > >& flat_domain_map,
+    Teuchos::RCP<const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node > >& flat_range_map,
     const LocalOrdinal block_size) {
     using Teuchos::ArrayRCP;
     using Teuchos::RCP;
     using Teuchos::rcp;
 
-    typedef Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node> Map;
-    typedef Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Node> Graph;
-
-    // Build domain map if necessary
-    if (flat_domain_map == Teuchos::null)
-      flat_domain_map = create_flat_map(*(graph.getDomainMap()), block_size);
-
-    // Build range map if necessary
-    if (flat_range_map == Teuchos::null)
-      flat_range_map = create_flat_map(*(graph.getRangeMap()), block_size);
-
-    // Build column map
-    RCP<const Map> flat_col_map =
-      create_flat_map(*(graph.getColMap()), block_size);
-
-    // Build row map if necessary
-    // Check if range_map == row_map, then we can use flat_range_map
-    // as the flattened row map
-    RCP<const Map> flat_row_map;
-    if (graph.getRangeMap() == graph.getRowMap())
-      flat_row_map = flat_range_map;
-    else
-      flat_row_map = create_flat_map(*(graph.getRowMap()), block_size);
-
-    // Build flattened row offsets and column indices
-    auto row_offsets = graph.getLocalRowPtrsHost();
-    auto col_indices = graph.getLocalIndicesHost();
-    const size_t num_row = graph.getLocalNumRows();
-    const size_t num_col_indices = col_indices.size();
-    ArrayRCP<size_t> flat_row_offsets(num_row*block_size+1);
-    ArrayRCP<LocalOrdinal> flat_col_indices(num_col_indices * block_size);
-    for (size_t row=0; row<num_row; ++row) {
-      const size_t row_beg = row_offsets[row];
-      const size_t row_end = row_offsets[row+1];
-      const size_t num_col = row_end - row_beg;
-      for (LocalOrdinal j=0; j<block_size; ++j) {
-        const size_t flat_row = row*block_size + j;
-        const size_t flat_row_beg = row_beg*block_size + j*num_col;
-        flat_row_offsets[flat_row] = flat_row_beg;
-        for (size_t entry=0; entry<num_col; ++entry) {
-          const LocalOrdinal col = col_indices[row_beg+entry];
-          const LocalOrdinal flat_col = col*block_size + j;
-          flat_col_indices[flat_row_beg+entry] = flat_col;
-        }
-      }
-    }
-    flat_row_offsets[num_row*block_size] = num_col_indices*block_size;
-
-    // Build flattened graph
-    RCP<Graph> flat_graph =
-      rcp(new Graph(flat_row_map, flat_col_map,
-                    flat_row_offsets, flat_col_indices));
-    flat_graph->fillComplete(flat_domain_map, flat_range_map);
-
-    return flat_graph;
-  }
-
-
-  // Create a flattened graph for a graph from a matrix with the
-  // MP::Vector scalar type (each block is an identity matrix)
-  // If flat_domain_map and/or flat_range_map are null, they will be computed,
-  // otherwise they will be used as-is.
-  template <typename LocalOrdinal, typename GlobalOrdinal, typename Device>
-  Teuchos::RCP< Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,
-                                 Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >
-  create_flat_mp_graph(
-    const Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> >& graph,
-    Teuchos::RCP<const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >& flat_domain_map,
-    Teuchos::RCP<const Tpetra::Map<LocalOrdinal,GlobalOrdinal,Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> > >& flat_range_map,
-    const LocalOrdinal block_size) {
-    using Teuchos::ArrayRCP;
-    using Teuchos::RCP;
-    using Teuchos::rcp;
-
-    typedef Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Device> Node;
     typedef Tpetra::Map<LocalOrdinal,GlobalOrdinal,Node> Map;
     typedef Tpetra::CrsGraph<LocalOrdinal,GlobalOrdinal,Node> Graph;
     typedef typename Graph::local_graph_device_type::row_map_type::non_const_type RowPtrs;