diff --git a/cmake/SetupBOUTThirdParty.cmake b/cmake/SetupBOUTThirdParty.cmake
index abe50913b2..d6d3f18b1e 100644
--- a/cmake/SetupBOUTThirdParty.cmake
+++ b/cmake/SetupBOUTThirdParty.cmake
@@ -25,6 +25,13 @@ if(BOUT_HAS_CUDA)
   # NOTE: CUDA inherits the CXX standard setting from the top-level
   # compile features, set for the bout++ target.
   set_source_files_properties(${BOUT_SOURCES_CXX} PROPERTIES LANGUAGE CUDA)
+
+  # timer.cxx contains no device code, and compiling its C++20 formatting and
+  # ranges machinery with cudafe++ is pathologically slow with CUDA 12.9.
+  set_source_files_properties(
+    ${PROJECT_SOURCE_DIR}/src/sys/timer.cxx PROPERTIES LANGUAGE CXX
+  )
+
   find_package(CUDAToolkit)
   set_target_properties(bout++ PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
   set_target_properties(bout++ PROPERTIES POSITION_INDEPENDENT_CODE ON)
diff --git a/include/bout/fieldops.hxx b/include/bout/fieldops.hxx
index 47cf6f8be8..a6f82a9970 100644
--- a/include/bout/fieldops.hxx
+++ b/include/bout/fieldops.hxx
@@ -474,7 +474,7 @@ struct BinaryExpr {
       this->div = div;
       return *this;
     }
-    BOUT_HOST_DEVICE BOUT_FORCEINLINE bool hasParallelSlices() const {
+    BOUT_FORCEINLINE bool hasParallelSlices() const {
       if constexpr (is_expr_constant_v<L> && is_expr_constant_v<R>) {
         return false;
       } else if constexpr (is_expr_constant_v<L>) {
@@ -485,7 +485,7 @@ struct BinaryExpr {
         return lhs.hasParallelSlices() && rhs.hasParallelSlices();
       }
     }
-    BOUT_HOST_DEVICE BOUT_FORCEINLINE int numberParallelSlices() const {
+    BOUT_FORCEINLINE int numberParallelSlices() const {
       if (!hasParallelSlices()) {
         return 0;
       }
diff --git a/src/mesh/parallel/shiftedmetric.cxx b/src/mesh/parallel/shiftedmetric.cxx
index ec8e19cc12..17b62257f4 100644
--- a/src/mesh/parallel/shiftedmetric.cxx
+++ b/src/mesh/parallel/shiftedmetric.cxx
@@ -440,7 +440,6 @@ static void shiftZ_block_fft(const int Nz, const BoutReal** in, BoutReal** out,
 
     fft_block_cooperative<32, FFTS_PER_BLOCK>
         <<<grid, block, 0, stream>>>(in, out, phs, nbatches, nblocks);
-
   } else if (Nz == 64) {
     constexpr int FFTS_PER_BLOCK = 4;
     constexpr int THREADS_PER_FFT = 64;
diff --git a/tests/unit/CMakeLists.txt b/tests/unit/CMakeLists.txt
index 60a5524c20..bb8a434761 100644
--- a/tests/unit/CMakeLists.txt
+++ b/tests/unit/CMakeLists.txt
@@ -67,7 +67,6 @@ set(serial_tests_source
     ./field/test_vector3d.cxx
     ./field/test_where.cxx
     ./include/bout/test_array.cxx
-    ./include/bout/test_rajalib.cxx
     ./include/bout/test_assert.cxx
     ./include/bout/test_bout_enum_class.cxx
     ./include/bout/test_boundary_iter.cxx
@@ -138,6 +137,10 @@ if(BOUT_HAS_HYPRE)
   list(APPEND serial_tests_source ./invert/laplace/test_laplace_hypre3d.cxx)
 endif()
 
+if(BOUT_HAS_RAJA)
+  list(APPEND serial_tests_source ./include/bout/test_rajalib.cxx)
+endif()
+
 add_executable(serial_tests ${serial_tests_source})
 
 target_include_directories(