diff --git a/cmake/MFCTargets.cmake b/cmake/MFCTargets.cmake
index e2ed8e736..39e4a4e0a 100644
--- a/cmake/MFCTargets.cmake
+++ b/cmake/MFCTargets.cmake
@@ -205,13 +205,16 @@ exit 0
                     # (a suffix of loop iterations silently skipped), which -O3 happens to
                     # mask today. Codegen is unchanged on amdflang; it is a win on upstream
                     # flang.
+                    # assume-no-thread-state: no kernel changes OpenMP ICVs on the device, so
+                    # the per-kernel thread-state bookkeeping is dropped (~1.2x on AFAR 24.3).
                     target_compile_options(${a_target} PRIVATE
                         -fopenmp
                         --offload-arch=gfx90a
                         -O3
                         -fopenmp-assume-threads-oversubscription
                         -fopenmp-assume-teams-oversubscription
-                        -fopenmp-assume-no-nested-parallelism)
+                        -fopenmp-assume-no-nested-parallelism
+                        -fopenmp-assume-no-thread-state)
                     # attributor-max-pi-accesses: amdflang generates device code for the WHOLE
                     # image at link time, and once the image carries enough target regions the
                     # device link's Attributor exceeds its AAPointerInfo access cap on a
diff --git a/src/simulation/m_global_parameters.fpp b/src/simulation/m_global_parameters.fpp
index e72afb668..583298952 100644
--- a/src/simulation/m_global_parameters.fpp
+++ b/src/simulation/m_global_parameters.fpp
@@ -9,7 +9,7 @@
 module m_global_parameters
 
 #ifdef MFC_MPI
-    use mpi  !< Message passing interface (MPI) module
+    use mpi
 #endif
 
     use m_derived_types